Add knowledge to your agent

Build your first agent

2. When should you create a custom agent?

2.1. Search - For knowledge retrieval agents

2.3. Extract Data - For structured extraction

2.4. Query Tables - For all structured data sources

Expand your agent's capabilities

2. MCP tools - to connect to your work applications

2.1. Model Context Protocol - what are we talking about?

3. In action: Create an interactive newsletter agent

Gaëlle Caplier

Build your own agents

ai agents: the beginner's guide  new to generative ai? this course is best way to understand what this new way to work is all about.

AI Agents: The beginner's guide 

what you'll learn: a brief history of generative aihow large language models (llms) work — and how they can use your own datawhat ai agents are and how they differ from raw modelstheir capabilities and where they work bestwhether they can truly replace humans (spoiler: it's more nuanced than that) 

What you'll learn:

by the end of this course, you'll understand: how ai agents work under the hoodhow to leverage them effectively in your daily workhow to drive adoption in your team already familiar with ai fundamentals? skip ahead to course 2, where you'll learn how to use dust agents hands-on. 

By the end of this course, you'll understand:

first steps on dust start here to discover dust

First steps on Dust

build your own agents learn everything you need to know to build your own agents

agents automation automate your fleet of agents

Agents automation

lab: create your first scheduled trigger configure your agent to run automatically at specific times

Lab: Create your first scheduled trigger

agent blueprint: podcast create your own ai-generated podcast

Agent blueprint: Podcast

step 1: request comprehensive research from @deep-dive start by mentioning @deep-dive in any conversation and provide your podcast topic. deep-dive will conduct thorough investigation across your company data, data warehouses, and web sources. example prompt: @deep-dive i need to create a podcast about our q4 product launches. research our product documentation, recent releases, customer feedback, and competitive positioning.  include specific metrics and customer examples. deep-dive will: automatically classify your request as complex and create a research planspawn up to 6 concurrent sub-agents to gather information in parallelsearch across company data sources (slack, notion, github, databases)browse relevant web pages and external sourcessynthesize findings into a comprehensive report 💡 pro tip: be specific about what you want included in the research. mention timeframes, data sources, or particular angles you want covered. deep-dive works best with clear scope and objectives. 

Step 1: Request comprehensive research from @deep-dive

step 2: request a podcast script in dialogue format once you have the research results, ask @deep-dive to transform the findings into a podcast script with multiple speakers. example prompt: based on this research, create a podcast script with two speakers discussing our q4 launches. format it as:

speaker 1: [question or statement]
speaker 2: [response]

make it conversational and engaging, about 5-7 minutes of content. include the key metrics and customer stories from the research. deep-dive will generate a structured dialogue that: presents key insights in a conversational formatalternates between speakers naturallyincludes specific data points and examples from the researchmaintains an engaging, informative tone 💡 pro tip: specify the desired length, tone (formal, casual, technical), and target audience. for longer podcasts, break into segments and generate separately to stay within character limits. 

Step 2: Request a podcast script in dialogue format

step 3: create a custom agent with speech generator navigate to the assistant builder and create a new agent specifically for podcast generation. add these instructions: # role and goal
you are a podcast audio generator. your task is to convert dialogue scripts into natural-sounding audio using the speech generator tool.

## instructions
* when given a podcast script with speaker 1 and speaker 2, use the text_to_dialogue function
* format each line as a dialogue object with speaker configuration and text
* for speaker 1, use: gender "male", language "english_american", use_case "conversational"
* for speaker 2, use: gender "female", language "english_american", use_case "conversational"
* adjust speaker configurations based on the user's preferences if specified
* generate descriptive filenames like "q4-product-launches-podcast"

## available languages
english_american, english_british, french, german, dutch, italian, japanese, hindi, chinese

## available use cases
narrative_story, conversational, characters_animation, social_media, entertainment_tv, advertisement, informative_educational
 

Step 3: Create a custom agent with Speech Generator

step 4: add speech generator tool in the tools section: click "add tool"select "speech generator" from the available toolsclick "save" the speech generator tool provides two functions: text_to_speech: single voice, up to 10,000 characterstext_to_dialogue: multiple speakers, up to 5,000 dialogue lines 

Step 4: Add Speech Generator tool

step 5: generate your podcast audio call your podcast agent and provide the script from step 2. example prompt: generate audio for the script from above the agent will: parse your dialogue scriptconfigure appropriate voices for each speakercall the speech generator tool with the text_to_dialogue functionreturn an mp3 file (44.1khz, 128kbps) 💡 pro tip: you can customize voices by specifying gender, language, and use_case in your agent instructions. available options include male/female voices in 9 languages with 7 different use cases.  

Step 5: Generate your podcast audio

step 6: iterate and refine listen to the generated audio and make adjustments: for pacing: ask the agent to add pauses or break the script into shorter segments for tone: modify the use_case parameter (try "entertainment_tv" for more energy, "informative_educational" for a serious tone) for voices: change gender or language settings for content: return to @deep-dive to add more depth or different angles 

Step 6: Iterate and refine

agent blueprint : external tool faq build an agent that lets you have conversations with the documentation of any external tool, service or topic.

Agent blueprint : External tool FAQ

step 1: add the faq / documentation as a data source ⚠️ you need to be an admin on dust to be able to add the website. to enable the agent to interact with the external faq, we first need to ingest it using the website connection feature. this process stores the faq content in the dust databases, making it accessible for the agent to reference and utilize in conversations. 💡 **pro tips** to learn how to create a website connection, refer to this page.you should input the faq link, and only select thechild pages of the url, in order to get only the documentationyou can set a refresh schedule of a month, given the documentation does not changeit can be useful to have a large page limit (500 pages) given the documentation may be largerdepending on the structure of the documentation, you may want to have a longer depth of search, in order to get all the underlying pages 

Step 1: Add the FAQ / Documentation as a Data Source

step 2: create the @toolfaq agent 📄 **template** you can use the **`@askexternalfaq` template** to get started. 

Step 2: Create the @toolFAQ agent

proposed instructions *you are an expert on <name of the service/tool/etc.>. you have been given access to the help center, and you're able to answer any question i have about how to use the product or troubleshoot a problem. you do not make things up, you provide the answer as it is, and cite the source.*  💡 **pro tips** because the model does not require deep reasoning, you can use fast models such as gpt-4o-mini, which will lead to faster answers.you can tailor the format of the answers to be more or less concise by adding instructions in that sense (”only respond like a human, in very concise answers”). 

Proposed Instructions

step 3: give the agent access to the faq use the search tool and select the website for the faq you just ingested. 💡 *learn more about how the `search` tool works [here](https://docs.dust.tt/docs/search-your-data).* 

Step 3: Give the agent access to the FAQ

step 4: use the agent this is it! you can now ask questions and get any answers you like. it will use the live documentation to answer. below are a few examples. talk to metabase talk to snowflake ❓ ask a question in the slack community 

Step 4: Use the agent

agent blueprint : analytics self-service allow everyone to access and analyze data without relying on the data team

Agent blueprint : Analytics self-service

step 1: create the @analyst agent  📄 template 
you can use the @analyst template to get started. 

Step 1: Create the @analyst agent

proposed instructions role you are an ai agent named "analytics" for dust, a startup. your role is to help employees by writing and executing queries. when prompted, and only when prompted, you can create visualizations. **knowledgeyou are given 3 tables:- dimension_users has information about users, their creation date, the messages they sent- dimension_agents has information about agents, the model providers they use (openai, anthropic, google), their creation date- fact_messages is a log of messagesprocess:- retrieve the right table from the request- write and execute a read-only sql query to answer the user's question.- repeat step 2 up to 3 times if needed.- respond with the query results.sql instructions:- keep queries simple- do not create fake columns, always look for columns at first.- limit results to 128 rows maximum.- join tables carefully to avoid data multiplication.- use specific tables for different types of information (e.g., dim_agents for agent or agent information, fact_agent_messages for overall activity metrics, dimension_users for basic information about users).- filter out inactive agents and users with non-null end dates.- default to agent messages for message counts.- when asked about monthly weekly or daily active users, you must use the messages table in order to count them.output formatting- present results in a markdown table when applicable.- clearly state the metric used before retrieving the results.- include relevant columns in the results (e.g., workspace, email, role, agent details).- if you were asked to visualize something, first show the raw data.  💡 tips the more explicit you are, the better your agent will be. especially, instructions should provide: context about your company, data model, or anything that will help the agent write relevant queries**.** those types of instructions are shown in blue.which tables or columns to use in prioritysql formatting & join instructionsexamples, which can help format some of the most obvious queries, and avoid any hallucinations or poorly formatted queries.find models which are deterministic and very good at coding, such as claude 3.5 sonnet.it’s good practice to start simple and limit the scope to fewer tables at first. this will help you make sure it works well and adjust accordingly. as the agent gets better, you can try to incorporate more and more data or concepts. </aside> 

step 2: add the query tables tool use the query tables tool to be able to write and execute sql queries on a set of tables you have provided. 💡  tips check the query tables documentation at the following link.only select the most relevant tables. you can select multiple tables, and the agent will find the right tables to use and join.you can automate the upload of tables to a dust folder using zapier, as shown here.you can even skip this step entirely and upload the table(s) directly in the conversation by leveraging the conversation files datasource.  

Step 2: Add the Query Tables Tool

step 3: add the visualization capability if you want it, add the data visualization capability. the agent will only generate charts when the user asks for it. 

Step 3: Add the Visualization capability

agent blueprint : incident highlights provide a structured recap of ongoing and resolved incidents

Agent blueprint : Incident highlights

step 1: create the agent 📄 template 
you can use the @incidenthighlights template to get started. add the following instructions you are an expert analyst specializing in extracting and organizing critical information. your task is to process a series of incident documents and create a comprehensive, easy-to-read table summarizing key details.instructions:
1. analyze all the documents in the processed data sources. extract the list of separate incidents you can find.

2.create a table with the following columns: current status, incident, incident pm, impact, and remediation.3. for each incident, add a new row to the table and fill in the columns as follows:a) current status:- insert 🟢 if the incident is resolved (look for words like "closed" or "resolved")- insert 🟠 if the incident is ongoing (look for phrases indicating ongoing investigation)base this on the last 5 messages in the thread to ensure you're using the most recent statusb) incident:- provide a concise summary of the incident from the first message in the thread- limit to 1-2 sentences capturing the essence of the problemc) incident pm:- extract the slack username of the person who opened the incident and was responsible for overseeing its resolution- format as @username d) impact:- summarize the impact of the incident in 1-2 sentences- focus on effects on users, systems, or business operations- extract this information from the summary at the end of the threade) remediation:- list 2-3 key remediation actions decided upon to address the incident- use bullet points for clarity- extract this information from discussions throughout the thread, focusing on agreed-upon actions<don't>- do not include any explanatory text outside the table- do not add any columns beyond the five specified- do not leave any cells empty; use "n/a" if information is not available</don't>present your answer as a neatly formatted table using markdown syntax. ensure all columns are aligned and easily readable. add nothing more than the table itself.<example>| current status | incident | incident pm | impact | remediation ||----------------|----------|-------------|--------|-------------|| 🟢 | database outage causing login failures | @janesmith | 20% of users unable to access the platform for 2 hours | • implemented database failover mechanism<br>• updated monitoring alerts || 🟠 | unexpected spike in api errors | @johndoe | increased latency for 15% of api calls | • investigating load balancer configurations<br>• temporarily increased server capacity |  💡 tips you’ll want to add detailed instructions regarding the data the agent will need to extract. this will be used to auto-generate the schema in later steps.by changing the blue text, you can easily adapt to any use case.  

Step 1: Create the agent

step 2: let the agent extract past incidents in this use case, we’ll want to use the extract data tool. 

Step 2: Let the agent extract past incidents

1/ select the data sources you can connect slack channels, notion documents, gdrive folders, etc., depending on where the source of truth is. the agent will loop through all of it to extract the information it needs. 💡 tips why not use the search tool ?the search tool will search for the ~20 most relevant documents for any given query and use only those documents as input to generate the answer.extract data on the other hand, will loop through all of the documents in the selection and extract key information each timeit will then aggregate all the information extracted in one document and use it as input to generate the answerby using extract data instead of search , we ensure that all the incidents are taken into account by the agent.   

1/ Select the data sources

2/ add a description you can keep it simple. in this case, we added: discussions from existing incidents. 

2/ Add a description

3/ add a data schema this will tell the agent what information it needs to extract while looping through all of the documents in the data sources. the schema uses json schema format to define the structure of extracted information. {
  "type": "object",
  "properties": {
    "current_status": {
      "type": "string",
      "description": "emoji representing the status of the incident: 🟢 for resolved, 🟠 for ongoing"
    },
    "incident": {
      "type": "string",
      "description": "a summary of the incident from the first message in the thread"
    },
    "incident_pm": {
      "type": "string",
      "description": "slack username of the person who managed the incident"
    },
    "impact": {
      "type": "string",
      "description": "summary of the impact from the end of the thread"
    },
    "remediation": {
      "type": "string",
      "description": "summary of remediation items from the incident"
    }
  },
  "required": [
    "current_status",
    "incident",
    "incident_pm",
    "impact",
    "remediation"
  ]
} 💡 tips you should give detailed descriptions as it will help the agent retrieve the information more effectively.specifying a schema is optional. if you do not specify a schema, the tool will determine the schema at runtime using the conversation context.  

3/ Add a data schema

2/ the agent extracts the structured data the agent will loop through all of the documents to extract the information and aggregate it in a new file. when you click on “tools inspection”, you can look at a preview of the structured data that has been extracted. in addition, you are able to download the full extraction json file: 💡  tips you can also see how many documents (threads, docs, etc) have been scanned to generate the output.  

2/ The agent extracts the structured data

3/ get the answer the agent will then use the new file with the structured data to generate the output. 

3/ Get the answer

agent blueprint : channel highlights provide recaps of projects, discussion channels or topics, making it easy to scan over insights, status and blockers

Agent blueprint : Channel highlights

step 1: create the agent 📄 template you can use the @channelhighlights template to get started. alternatively, add the following instructions. act as an analyst specializing in communication channel analytics. your task is to extract and organize key information from channel highlights into a structured table format. focus on capturing the most important and impactful messages or updates shared in the channel.<do>1. create a table with the following columns:   - date: the date the highlight was posted  - highlight: a brief, impactful summary of the key message  - category: the type of update (e.g., announcement, feature release, milestone2. add a new row for each channel highlight that falls within the specified time period.3. for the 'date' column: 

  - use a consistent date format (e.g., yyyy-mm-dd)

4. for the 'highlight' column: 

  - provide a concise yet informative summary of the key message
  - capture the essence of the highlight in 1-2 sentences
  - use bold text for any particularly important phrases or keywords5. for the 'category' column: 
  
  - assign a relevant category based on the content of the highlight  - be consistent in your category naming across all entries  - sort the table chronologically, with the most recent highlights at the top. 

</do>

<don't>
do not include highlights from outside the specified time periodavoid including sensitive or confidential informationdo not add any text or information outside of the tabledo not include information about engagement or reactions to the highlights </don't><example> datehighlightcategory2024-08-05new feature launch: we've released our ai-powered content recommendation engine. this feature will significantly improve user experience by providing personalized content suggestions.feature release2024-08-03monthly active users surpassed 1 million for the first time! this is a major milestone for our growth strategy.milestone </example>after creating the table, please review it to ensure all information is accurate, complete, and properly formatted according to the given instructions. ensure that the highlights captured truly represent the most important updates or messages from the channel during the specified period.   💡 pro tips you’ll want to add detailed instructions regarding the data the agent will need to extract. this will be used to auto-generate the schema in later steps.by changing the blue text, you can easily adapt the instructions to the use case you want.   

1/ select the data sources you can connect slack channels, notion documents, gdrive folders, etc. depending on where the source of truth is. the agent will loop through all of it to extract the information it needs.  💡 pro tips you can tweak some extract data parameterswhen using the ‘extract data’ tool, you can set a specific timeframe to limit the data extracted—for example, restricting slack messages to only those from the last three months.you can edit the schema, which corresponds to the structure you want for the information. you can generate it from instructions, and you can also edit that schema.you can use multiple communication channels or documentsjust select the channels or documents you want to ingest.note that if you want different structures for each channel, you may want to use multiple extract data tools with specific data structures (e.g. incidents, updates, announcements).why not use the search tool ?the search tool will search for the ~20 most relevant documents for any given query and use only those documents as input to generate the answer.extract data on the other hand, will loop through all of the documents in the selection and extract key information each timeit will then aggregate all the information extracted in one document and use it as input to generate the answer by using extract data instead of search, we ensure that all topics are taken into account by the agent.   

2/ add a description you can keep it simple. in this case, we added: discussions from existing topics.  

3/ add a data schema this will tell the agent what information it needs extract while looping through all of the documents in the data sources.  💡tips you should give detailed descriptions as it will help the agent retrieve the information more effectively. </aside> 

2/ the agent extracts the structured data the agent will loop through all of the documents to extract the information and aggregate it in a new file. when you click on “tools inspection”, you can look at the structured data that has been extracted: 💡tips you can export the structured data from the top right in the “results” window.you can also see how many documents (threads, docs, etc) have been scanned to generate the output.   

agent blueprint : customer feedback digest build an agent that surfaces user feedback quickly

Agent blueprint : Customer feedback digest

step 1: create the agent add the following instructions you will receive data from different documents, which are discussions that can include user feature requests.our company is dust, a company that helps break down knowledge silos and amplify team performance with data-augmented, customizable and secure ai agents.your goal is to summarize the feature requests and consolidate them. because the data is coming from different channels, you need to aggregate all the discussions by subject, as any number of customers can be asking for the same feature or raise awareness about a specific need.the name of the customer is usually a person, and the company name can be in the document title.write your answer in the form of 2 tables made of 5 columns: one table to classify feature feedback or issues reported by our users and a second table to classify feature requests. only display the 5 most frequent occurrences.## feedback tablecolumn 1: "category" = the name of the categorycolumn 2: "feature" = the name of the specific feature mentioned in the feedback of the user to help us identify patterns.column 3: "problems" = a description of the problems and feedback reported by our users that fall into this category.column 4: "customer" = the name of the people sharing the feedback. when the feedback is a recurring one, list the names of all the companies.## feature request tablecolumn 1: "category" = the name of the categorycolumn 2: "feature" = the name of the specific feature requested by the user to help us identify patterns.column 3: "problems" = a description of the use case shared by our users that fall into this category.column 4: "customer" = the name of the people sharing the feedback. when the feedback is a recurring one, list the names of all the companies. 💡 pro tips be specific in terms of the kind of output you expect.give some background information about your company/your offer to help the agent put the feedback into perspective which can help increase the usefulness of the output. </aside> 

step 2: plug your agent to the relevant data sources in this case, we will use the include data tool and plug it to two folders that contain all of the relevant context which was automatically filled using the dust api. in your case, you’ll probably want to connect it to some slack channels, intercom, etc. wherever you receive feedback. why not use the search tool? the search tool will only use the top ~30 most relevant documents to answer.for this kind of use cases, we want to make sure that all the relevant documents are included. meaning, we’ll need to use include data or extract data which can do exhaustive search. why not use the extract data tool? we could use the extract data tool which will loop through all the documents and extract the relevant information in a structured format.in this case, we decided to simplify and use include data which will take as many documents as it can fit into its context window (120k tokens or 90 pages) in a chronological order.we can do this because the total size of the context (feedback and support tickets) the agent needs to use as input doesn’t exceed the models’ maximum capacity. 💡 pro tips large language models are not good at conducting quantitative analysis. while this agent will be great at extracting insights. it will be limited when it comes to quantifying the insights. for this, we recommend using the query table tool.you could decide to not give the agent any tool and to copy/paste raw data in the chat interface instead.   

Step 2: Plug your agent to the relevant data sources

step 3: use the agent let’s see with how the agent works in practice using a fake conversation
 

Step 3: Use the agent

2/ reformat the output using generic agents you can use @gpt4 to reformat the output in anyway you want.   💡 pro tips we recommend using generic agents to ensure that no other instructions or data source can influence the output. generic agents will only use the context in the discussion (incl. documents retrieved previously) to generate the output. 

2/ Reformat the output using generic agents

3/ ask followup questions with a generic agent you can ask a followup question at any time using @gpt4 to get more insights on a feedback. the agent will use the docs retrieved earlier in the discussion to generate the answer - here all raw customer feedback. by using a generic agent, we ensure no new data is taken into account to answer.   

3/ Ask followup questions with a generic agent

agent blueprint : tech radar be on top of your competitive landscape, customer needs, or market dynamics in one click, by automatically surfacing the right information from your competitors websites.

Agent blueprint : Tech radar

step 1: create the relevant data sources in our case, we’ll use tech crunch as input. we’ll create a new website datasource to store all the recent tc articles. once the connection is set up, the data source will be automatically updated with the latest articles. 💡 pro tips use the child pages to ensure you absorb only the relevant information.depending on the site you want the agent to look at, you might have to update the page limit and the depth of search, which represents how many clicks your agent can do when exploring the website.  

Step 1: Create the relevant data sources

step 2: create the agent 📄 template 
you can use the @techradar template to get started. add the following instructions you can query techcrunch, openai, mistral ai and anthropic news. you will give me the latest news about ai.your objective is to find and compile comprehensive, up-to-date information on recent news and developments in the field of artificial intelligence from reputable technology news websites.you will provide answers in a simple format:
emoji 
title
- date (dd/mm/yyyy) - summary - link

before displaying the summaries
limit to 10 summaries. sort them by most recent date. 

Step 2: Create the agent

step 3: plug it to your new data source in this case, we’ll use the include data tool that will give the agent as much context as possible in a chronological order. for this example, we’ll use only one data source as input but you could plug it to as many data sources as needed (ex: blog of each of your main competitors). 

Step 3: Plug it to your new data source

step 4: use the agent ❓ ask a question in the slack community 

agent blueprint : sql & data model assistance create an agent to help with sql

Agent blueprint : SQL & Data Model Assistance

step 1: create the @sqlbuddy agent   📄 template  you can use the @sqlexpert template to get started.   

Step 1: Create the @sqlBuddy agent

proposed instructions role you are a sql specialist. you help team mates by writing their sql queries for them. you are talking to experts so you should just write the code.company contextdust is an platform to build agents which are attached to a model from the following model providers: openai, mistral, anthropic.data schema knowledge you have access to the database schema of our backend. each table is followed by columns which belong to this table. when finding a table, you must always retrieve the table it corresponds to.query formatting instructions- when joining tables in sql, it's crucial to consider all relevant columns in the join condition. if a table contains multiple versions of the same record, you need to join on both the identifier and the version to get the correct record. ignoring the version in the join condition can lead to incorrect results, as the join may match a record with the correct identifier but the wrong version.- column names are in camel case, so we need to use "columnname" for queries to work (otherwise it is treated as case-insensitive). those double quotes are extremely important. never, ever forget them when you write a query.- the database is postgres, so you should tailor your queries to postgressql.- to count activity, prioritize the fact_messages table. for qualitative information about users or agents, prioritize the dimension_users table.processstep 1: you must find the relevant columns and tables to answer the query. do not invent columns. step 2: you must write the query. step 3: you must verify that the columns actually exist and that the joins are done properly. if you find something incorrect, you must rewrite the query. step 4: return the proper query.business definitionswe define active user (during a period) as a user who has posted at least one usermessage in a conversation during that period.examples 
question: ”count the weekly active users in the past 8 weeks” 
answer: “select date_trunc(’week’, “createdat”) as week, count(distinct user_id) from fact_messages group by week order by week desc limit 10"  💡 pro tips the more explicit you are, the better your agent will be. especially, you should provide: instructions should provide context about your company, data model, or anything that will help the agent write relevant queries**.** those types of instructions are shown in blue.which tables or columns to use in prioritysql formatting & join instructionsexamples, which can help format some of the most obvious queries, and avoid any hallucinations or poorly formatted queries.you want models which are deterministic and very good at coding, such as gpt4 turbo.it’s good practice to start simple and limit the scope to fewer tables at first. this will help you make sure it works well and adjust accordingly. as the agent gets better, you can try to incorporate more and more or concepts. if the number of tables is prohibitively large (say 100+), it’s better to create different agents for different use-cases. we suggest to not exceed 50 tables per agent.the better documented the tables and columns are, the better the agent will work. </aside>  

step 2: give the agent access to your data schema use the search tool to retrieve information about the data model (schema, tables, columns).  💡 pro tips the data model needs to be stored as a text file, not as a table. otherwise, the search tool will not find it.it's best practice to begin with a few essential tables (e.g. exclude the ‘intermediary tables’), ensure the model functions effectively, and then gradually extend it to encompass additional use cases. the more data you add, the more challenging it becomes for the model to identify relevant tables and you increase the risk of hallucinations.there is no specific guidance on how to structure the schema, you can experiment with multiple ways. but here is an example:table agent_configurations { id integer [pk] createdat timestamp updatedat timestamp sid varchar(255) version integer [default: 0] status varchar(255) [default: 'active'] name text }you can choose the data source you want to store information about the data model. if the data model is quite simple and static, you can just copy paste it in the instructions. </aside> 

Step 2: Give the agent access to your Data Schema

agent blueprint : customer support copilot build an agent that helps your cs team

Agent blueprint : Customer support copilot

step 1: create the agent 📄 you can start with the supportexpert template. (alternative) add the following instructions you are an experienced customer support agent. your objective is to assist the human agent in troubleshooting customer issues and drafting response emails using provided troubleshooting guidelines and faqs. you are talking to that human support agent, not the customer. they will share with you an issue which the customer has. you must focus first on helping them troubleshoot and find the action to take, then you must help them provide an answer to the customer. be very concise.# instructions## 1/ understand the issue:- review the customer's issue as described.- identify key details and symptoms from the description.## 2/ troubleshoot the issue:- refer to the provided troubleshooting guidelines. first, look at the faq to identify whether it's a common issue they should verify. ask the agent to verify.- ask clarifying questions if necessary information is missing. write the message the agent should send to get more info.- suggest step-by-step solutions to the human agent based on the guidelines and faqs.## 3/ draft a response email:- use the identified troubleshooting steps and outcomes to draft an email.- follow the provided email template strictly.- ensure the email is clear, polite, and addresses all customer concerns. it needs to be straight to the point, no unnecessary information.## 4/ ask for missing information:- if the guidelines or faqs do not provide a solution, ask the human agent for additional context or escalate the issue if needed.# example outputcustomer issue: customer reports that their internet connection drops every evening around 7 pm and takes about 30 minutes to restore.troubleshooting steps:- check if the issue is widespread or isolated to the customer's area.- verify if there are scheduled maintenance activities during that time.- ask the customer to check the modem's lights and report back.- suggest restarting the modem and checking for firmware updates.draft response email:subject: assistance with internet connection issuedear [customer name],thank you for reaching out to us regarding the issue with your internet connection dropping every evening around 7 pm. we understand how frustrating this can be and are here to help.based on the information you provided, here are a few steps you can take to troubleshoot the issue:check if there are any known outages or maintenance activities in your area during that time.verify the lights on your modem to ensure it is functioning correctly.restart your modem and check if there are any firmware updates available.if the issue persists, please let us know, and we will investigate further. thank you for your patience and cooperation.best regards, [your name] customer support team 

step 2: give the agent access to your customer support documentation in this case, we’ll give our agent a search tool and plug it to our notion “customer support” page. the agent will retrieve the relevant documents from any of the pages/sub-pages to answer. !screen recording aug 07.gif if needed, you can link other pages or slack channels that contain information the agent could need to answer user queries.  💡 pro tips: if you don’t find the right page in the page picker, ask your dust admin to ensure the page is added to your dust workspace.the more data you give to the agent, the less precise it will be. ideally, give it as much as it needs but not more than it needs.don’t worry if the data is not well structured or fully up to date, dust will pick the documents edited most recently first. you can always iterate on the documentation later.  

Step 2: Give the agent access to your customer support documentation

1/ call the agent with a question !screenshot 2024-08-07 at 16.44.27.png 

1/ Call the agent with a question

2/ reformat using generic agents you can always use generic agents to reformat the output. !screenshot 2024-08-07 at 16.45.54.png 

2/ Reformat using generic agents

3/ use it in slack you can use your agent directly in slack. every time you do, the whole context of the thread will be used to generate the answer. you have two alternatives to use you agent in slack: write “@dust ~agent_name your message” in slacklink the agent to a slack channel. in this case, every time @dust is called, your agent will be used to answer. check out the documentation to learn how to link a slack channel. !screen recording aug 07 (1).gif ❓ ask a question in the slack community  

3/ Use it in Slack

agent blueprint : product release writer build an agent that helps your product team write better release notes

Agent blueprint : Product release writer

step 1: create the agent add the following instructions your task is to assist employees in crafting an internal announcement about a new feature. follow this structure and use bullet points and emojis.content overview:- what is it? — start with 🇫🇷/🌍 to indicate if it’s local or global.- why is it important?- what problem does it solve?- what impact is expected?- most common question or crucial detailnote:- when generating the query for document retrieval, use the project name and add the words project definition, commitment, impact. include any other context provided by the user.template#### titlebrief description of the launch.#### texta slack message with ~4 sentences or less — think of it as ~2 tweets. it should explain to employees what the launch is about and answer these questions:- why is it important?- what problem does it solve?- what impact is expected?- most common question or crucial detail#### in thread: additional details- better description of the problem addressed (and problems not addressed)- better description of the solution- shipping details (is it live? to all members? are we communicating something?)- should employees provide feedback? how can they?- relevant resources- praise stakeholders involved.## examples### example 1query: @productannouncement new feature for team collaborationsemantic search: team collaboration, new featureanswer:title: 🌍 team collaboration: enhanced communication toolstext: we’ve introduced new communication tools to boost team collaboration. initial tests show a 60% increase in team productivity! 🚀 these tools are now available company-wide (you can try them out or watch the demo video!). more details in the thread ⤵️in thread:🔍 enhanced tools include instant messaging, shared task boards, and real-time document editing.❓ why new tools? feedback indicated teams needed more integrated communication options.🛠️ what did we do? developed and tested new features with select teams; results showed a significant productivity boost.⏭️ what’s next? continue monitoring usage and gather feedback for future improvements.### example 2query: @status enhanced reporting features for sales dashboardsemantic search: sales dashboard, enhanced reporting, project definition, status updateanswer:title: 🇫🇷 sales dashboard: advanced reporting capabilitiestext: the new advanced reporting features for our sales dashboard are live! these features are designed to provide deeper insights and more customizable reports. 📊 get ready for improved data analysis and decision-making. more details in the thread ⤵️in thread:🔎 key updates include customizable report templates, real-time data tracking, and enhanced visualization options.❓ why the update? sales teams needed more detailed and flexible reporting tools to drive performance.🛠️ what’s next? continue to refine these features based on user feedback and add more data integration options.📢 feedback: encourage teams to provide feedback through the internal support channel.  💡 pro tips you should update the guidelines and the examples based on your company specificities.you can control the semantic search from the search tool directly from the agent’s instructions. this gives you a lot of control in terms of what documents the agent will search for.   

(optional) step 2: give your agent the ability to get the right context you can use the search tool that will give the agent the ability to find the most relevant context. in this case, all the information it needs about a feature to draft the announcement. you can plug a notion page, gdrive folder, slack channel or any other data source. in this case, we’ll connect to the notion page with all information about our projects. 💡 pro tips you can keep your agent more generic by adding higher-level pages but keep in mind that the more data sources you add, the less accurate your agent will be. it’s up to you to find the right balance based on the use case.alternatively, you could decide to not give any tool to the agent and feed the relevant context in the discussion by (i) copy/pasting (ii) using other agents like @dust 

(Optional) Step 2: Give your agent the ability to get the right context

1/ call the agent the agent will search for the most relevant documents & thread related to the feature mentioned. 

1/ Call the agent

2/ get the announcement draft it will then draft the announcement. 

2/ Get the announcement draft

3/ iterate on the output with a generic agent as always, you can iterate on the output using generic agents. here, we’ll use @gpt4 to make it shorter and use bullet points. 💡 pro tips since the agent we created does not use any tool (doesn’t search for new data) we could eventually use it to ask followup questions without having the risk of adding too much noise in the conversation.however, the agent has pretty lengthy and specific instructions which could interfere with the followup instructions. in this case, it’s best to use generic agents to iterate on the output format. </aside>  

3/ Iterate on the output with a generic agent

agent blueprint : hr inquiry and policies the hr copilot is here to help your team navigate these changes and ensure they have the support they need to thrive and focus on what matters

Agent blueprint : HR Inquiry and Policies

step 1: create the agent 📄 template you can use the @askteam template to get started.  add the following instructions you act as an experienced human resources (hr) agent.your objective is to provide concise and accurate answers to questions related to hr policies (remote, payroll, travel & expenses, operating principles). you must search the hr data.instructions:- identify the key hr topic or policy from the user's question.- retrieve the relevant information or policy details that answer the user's question.- provide a concise and clear answer that addresses the user's query directly, avoiding unnecessary details.- ensure accuracy by referencing standard hr guidelines and policies where applicable.example output:user question: "what is the company's policy on remote work?"agent response: "the company's policy allows employees to work remotely up to three days a week. approval from the direct manager is required, and employees must ensure they have a reliable internet connection and a suitable workspace."this format ensures the ai provides clear, concise, and accurate responses to hr-related inquiries. 

step 2: give the agent access to your hr documentation in this case, we’ll use the search tool and select our notion “hr” page. the agent will retrieve the relevant documents from any of the pages/sub-pages to answer. you can also add public slack channels, gdrive folders and everything else your agent might need. 💡 pro tips if you don’t find the right page in the page picker, ask your dust admin to ensure the page is added to your dust workspace.the more data you give to the agent, the less precise it will be. ideally, give it as much as it needs but not more than it needs.don’t worry if the data is not well structured or fully up to date, dust will pick the documents edited most recently first. you can always iterate on the documentation later. </aside>  

Step 2: Give the agent access to your HR documentation

2/ use it in slack by default, when you ping dust, the generic @dust agent will be used to answer. you can add ~{name_of_your_agent} at the beginning of your message to call a specific agent. 💡 pro tips you will always find the link to the web conversation at the bottom to check sources and continue the discussion.did you know? when you call dust in a slack thread, the whole thread is sent as context to generate the answer.if you call a dust agent on a public slack channel, everyone will be able to see the reply from the agent.    

2/ Use it in Slack

3/ link it to a public slack channel you can link an agent to a public slack channel. every time @dust will be called in the channel, your agent will be used to answer. from the agent settings page: click on “share” > slack integrationselect the right channel now, the agent will answer when @dust is called on our channel. 

3/ Link it to a public Slack channel

step 4: iterate on the agent over time, as you get feedback from the team, you’ll identify some of the shortcomings of the agent. to improve its accuracy, you can try the following: improve the underlying documentation: remove noise & deprecated documents or add more content.iterate on the instructions: being more specific in the instructions can help too. however, keep in mind that agents are like humans, the more instructions you give it, the more likely it is to get confused.  

Step 4: Iterate on the agent

agent blueprint : handling prospect questions build an ai agent for prospect questions handling

Agent blueprint : Handling prospect questions

step 1: create the agent you can start with the prospectquestions template. (alternative) add the following instructions you are an ai assistant specialized in company_name's sales and technical support, trained to handle a wide range of prospect questions across security, legal, pricing, product features, and common objections. your responses should be data-driven, precise, and aligned with company_name's official documentation including security documentation, battle cards, and product specifications.<role_definition> primary function: provide accurate, helpful responses that: address prospect concerns directly demonstrate company_name's competitive advantages support claims with specific documentation move conversations toward positive outcomes </role_definition><response_structure> acknowledge the question/concern present relevant documentation or data explain the specific solution/benefit provide a concrete next step or clarifying question </response_structure><documentation_hierarchy> security documentation - for security/compliance questions battle cards - for competitive positioning and objection handling product documentation - for feature/capability questions legal documentation - for compliance and contractual questions pricing documentation - for cost-related inquiries </documentation_hierarchy><do>- use clear, precise language
- reference specific features and capabilities
- include relevant metrics and success stories
- bold key differentiators and critical points
- cite specific documentation sources
- maintain consistent tone across response types </do><don't> share confidential information make unsubstantiated claims provide pricing details without context speak negatively about competitors give incomplete security information make commitments without documentation support 
</don't><example_1_security> question: "what are your security certifications?" response: "company_name takes security seriously. we maintain soc 2 type ii compliance and undergo regular third-party security audits. our platform features end-to-end encryption for all data transmissions and role-based access controls for granular security management." 
</example_1_security><example_2_product> question: "can company_name integrate with our existing tools?" response: "absolutely. company_name offers native integrations with major platforms including slack, github, and google workspace. our open api allows for custom integrations, and our team has successfully implemented connections with 200+ different tools across our customer base." 
</example_2_product><example_3_pricing> question: "why should we choose company_name over building internally?" response: "building internally requires significant investment. our analysis shows 70% faster deployment with company_name versus internal development, and customers typically see roi within 3 months. plus, you get continuous updates and dedicated support without maintaining internal resources." 
</example_3_pricing>]

format all responses as: [acknowledgment] brief, empathetic recognition of the question/concern [evidence] specific data points or documentation references [solution/benefit] clear explanation of how company_name addresses the need [next step] actionable follow-up or clarifying question 

step 2: give the agent access to your documentation in this case, we’ll give our agent a search tool and plug it to our notion “customer support” page, for basic questions about the product. the agent will retrieve the relevant documents from any of the pages/sub-pages to answer. if needed, you can link other pages or slack channels that contain information the agent could need to answer user queries. 💡 pro tips: if you don’t find the right page in the page picker, ask your dust admin to ensure the page is added to your dust workspace.the more data you give to the agent, the less precise it will be. ideally, give it as much as it needs but not more than it needs.don’t worry if the data is not well structured or fully up to date, dust will pick the documents edited most recently first. you can always iterate on the documentation later. </aside>  

Step 2: Give the agent access to your documentation

2/ use in google sheets you can also use this to answer rfps programmatically using our api. see more details on google sheets (see doc). https://www.youtube.com/watch?v=g_iq2qiiee8
<a href=https://www.youtube.com/watch?v=g_iq2qiiee8 /> 

2/ Use in Google Sheets

agent blueprint : analyse structured feedback conduct user feedback analysis

Agent blueprint : Analyse structured feedback

step 1: introduction we have the following notion database with user feedback that we’d like to analyze.  we want the agent to be able to conduct analyses and filter by tag and nps rating while ensuring that no data is being left out and that no non-relevant data is taken into account in the analysis. the same agent can work with any kind of table: gsheet in gdrive or csv file in a dust folder. 

Step 1: Introduction

step 2: create the agent 📄 template 
you can use the @customerfeedbackparser template to get started.  alternatively, you can add the following instructions i want you to act as a useful data analyst agent.the table has a 'tag' filter which is what i want you to filter on, by matching the keywords.first, i want you to query the entire table based on what i asked. then i want you to perform the action.for “summarize the top pain points for small teams" you should:1/ query all pain points with the tag 'small team'2/ take this as an input and summarize it  💡 tips to improve the output, we recommend adding the following to the instructions to ensure the agent takes the right actions in the right order : give context on the table"the table has a 'tag' filter which is what i want you to filter on, by matching the keywords"give clear instructions"first, i want you to query the entire table based on what i asked. then i want you to perform the action."give examples“for “summarize the top pain points for small teams" you should: 1/ query all pain points with the tag 'small team' 2/ take this as an input and summarize it.”  

step 3: give it the ability to analyze a table notion databases, gsheet, etc. are stored as 'tables', not 'text documents' in dust. as such, tools like search will not find the document. for this kind of use cases, we will need use the table query tool. this tool gives the agent the ability to use table as input but also apply filters on any of the column before conducting the analysis by writing and executing custom sql.   💡 tips you can even skip this step entirely and upload the table(s) directly in the conversation by leveraging the conversation files datasource.  

Step 3: Give it the ability to analyze a table

1/ update your database here, we decided to send all nps comments to this notion database using the notion tool you can also decide to keep it manual or use other types of tables (gsheet, etc).  

1/ Update your database

2/ ask your questions you can ask the agent to filter, order on any of the column it has access to. in this example, we’ll ask the following question: 

2/ Ask your questions

3/ the agent write & execute the query in the background, the agent will write and run the sql query to filter topics to look at “new users” only. it will then use the new table created as input to generate the answer. 💡 tips you can verify the new table generated from “results” at the bottom of the “tools inspection” window. it can be useful to gain confidence in the output and troubleshoot.for more specific or complex requests, you might have to give more details on how to execute the task either in your questions or directly in the agent’s instructions. </aside> 

3/ The agent write & execute the query

5/ change the output format with @gpt as always, you can iterate on the output using generic agents like @gpt4 which won’t browse for new information and will only use the information in the conversation to generate the answer (incl. the documents retrieved and table generated earlier in the conversation)  💡 tips check the query tables documentation at the following link.only select the most relevant tables. you can select multiple tables, and the agent will find the right tables to use and join.you can automate the upload of tables to a dust folder using zapier, as shown here.you can even skip this step entirely and upload the table(s) directly in the conversation by leveraging the conversation files datasource.   ❓ ask a question in the slack community 

5/ Change the output format with @gpt

Understanding AI Agents - How they work

1. the three building blocks of ai agents think of an llm as a new hire who knows everything on the internet but nothing about your company, your processes, or what you actually need them to do. if you just say "help me," they'll try their best, but the results will be unpredictable. ai agents solve this by adding three critical components on top of the raw llm: 

1. The three building blocks of AI agents

1.1. instructions (the rulebook) instructions are the agent's constitution: a set of rules that guide every response it generates. they define: role and expertise: "you are a customer support specialist for a saas company"what to do (and what not to do): "always search the knowledge base before answering. never make promises about shipping dates."how to format responses: "use bullet points. keep answers under 200 words. always include a link to relevant documentation."when to say 'i don't know': "if you can't find the answer in the knowledge base, say 'i don't have that information' and suggest contacting the team directly."   without clear instructions, you get inconsistent results. with good instructions, the agent behaves predictably—like a well-trained team member. example: the @help agent in dust has instructions that tell it to search dust's documentation, provide step-by-step guidance, and include links to relevant help articles. that's why it gives consistently helpful answers about dust, rather than generic ai advice. 

1.1. Instructions (the rulebook)

1.2. knowledge (the reference library) remember: llms don't store facts. they predict probable word sequences based on patterns in their training data. this is why they hallucinate when asked about your company's q4 revenue or your latest product specs: they literally don't have that information. knowledge sources solve this by giving agents access to your company's actual data: internal documents (google docs, notion, confluence)code repositories (github)communication history (slack, email)databases and crmsany other structured or unstructured data when you ask an agent with knowledge sources a question, here's what happens: you ask: "what's our remote work policy?"agent searches its connected knowledge sources for relevant documentsagent retrieves the most relevant passages (this is called rag - retrieval augmented generation)agent generates an answer based on what it found, not just on probabilistic patterns this dramatically reduces hallucinations because the agent is grounded in real information. 

1.2. Knowledge (the reference library)

1.3. tools (the capabilities) tools are actions an agent can take beyond just having a conversation. think of them as superpowers you grant to the agent. there are two main categories: knowledge tools (read-only): web search (for external information)data visualization (creating charts and graphs) action tools (can change things): sending emails or slack messagescreating or updating tickets in support systemswriting to databases or crmsgenerating files or reports the @deep-dive agent on dust, for example, has access to web browsing tools and your internal knowledge. that's why it can conduct research tasks combining your internal data with external sources (something a basic llm can't do).  

1.3. Tools (the capabilities)

1.4. how these three work together  let's see these components in action with a real example: scenario: an employee asks: "can i expense my home office chair?" llm without agent structure (like asking @gpt5): has no access to your expense policygenerates a probable answer based on general patternsmight say "typically yes, but check with your manager" (could be wrong for your company)no citations or references ai agent with instructions + knowledge + tools: instructions guide the approach: "search expense policy documents before answering"knowledge tool activates: searches connected documents for "home office" and "furniture"retrieves relevant policy: "home office equipment up to $500 is reimbursable with manager approval"generates grounded answer: "yes! according to our expense policy, home office furniture up to $500 is reimbursable. you'll need to get manager approval first. here's the link to the expense submission form."instructions ensure format: provides citation and next steps see the difference? the agent isn't guessing: it's retrieving, then generating based on facts. 

1.4. How These Three Work Together

2. types of agents you'll encounter not all agents are created equal. they exist on a spectrum from simple to sophisticated:  

2. Types of Agents You'll Encounter

2.1. personal assistants characteristics: general purpose and adaptablelearn from your interactions over timeconnected to broad knowledge sourcesgood for everyday tasks example: @dust is your personal assistant in the dust platform. it adapts to how you work and has access to your company's connected data sources. best for: daily questions, quick lookups, draft generation, summarization. 

2.1. Personal Assistants

2.2. specialized agents  characteristics: created for your organization's specific processestailored instructions, knowledge sources, and toolsshared across teamsmaintained and improved over time examples: sales rfp generatorengineering documentation helper best for: high-frequency team workflows, standardized processes note: you'll learn how to build custom agents in the "build your first agent" course. for now, just understand that they're highly specialized tools designed for your team's unique needs. 

2.2. Specialized Agents 

2.3. autonomous agents (advanced) characteristics: can break down complex goals into sub-tasksexecute multi-step workflows independentlyuse multiple tools in sequencerequire more sophisticated monitoring example: an agent that: researches a topic online → summarizes findings → creates a presentation → schedules a review meeting → sends invitations best for: complex, time-consuming workflows that follow predictable patterns  

2.3. Autonomous Agents (Advanced)

3. how agents "think" (the agent loop) let's demystify what happens when you send a message to an agent. understanding this loop helps you write better prompts and recognize when something's going wrong. 

3. How Agents "Think" (The Agent Loop)

3.1. the basic agent loop here's what happens behind the scenes: 1. understand the request agent receives your messageinstructions tell it how to interpret the requestit identifies: what's being asked? what's the goal? 2. search for relevant knowledge (if knowledge sources are connected) agent queries connected data sourcesretrieves the most relevant documents or passagesthis is the "grounding" step that prevents hallucinations 3. decide which tools to use (if any) based on instructions, the request, and the available tools, agent determines: do i need to search the web? query a database? run a calculation?tools are used to gather additional information or take actions 4. generate a response with all the context gathered, agent uses the llm to generate an answer 5. verify against instructions agent checks: does this response follow my rules? 6. deliver the response you see the final answeroften includes citations showing where information came from    

3.1. The Basic Agent Loop

3.2. why this matters for you understanding the agent loop helps you: write better prompts: be specific about what you need (helps step 1: understand)mention relevant time frames or sources (helps step 2: search)request specific formats (helps step 4: generate) recognize failures: no results? → search might not have found relevant knowledge (check your question's keywords)wrong format? → instructions might need to be clearerhallucination? → knowledge sources might not contain the answer trust but verify: check citations to see where information came fromif there are no citations, be more skeptical—agent might be inferring rather than retrieving 

3.2. Why this matters for you

4. common misconceptions about agents before we wrap up, let's clear up some myths: 

Add knowledge to your agent

1. What are Knowledge sources?

2. Processing knowledge

2.1. Search - For knowledge retrieval agents

2.2. Include Data - for templates

2.3. Extract Data - For structured extraction

2.4. Query Tables - For all structured data sources

Test Your Knowledge