Introducing: multi-tool assistants

Dust assistants can now pick and choose from several tools, thinking in steps in order to best answer user queries. We’ve also expanded the list of tools that can be used. We believe this will help our users enhance an ever-growing set of workflows in their day-to-day.

Pauline Pham

• 5 min read
Introducing: multi-tool assistants

Dust assistants can now pick and choose from several tools, thinking in steps in order to best answer user queries. We’ve also expanded the list of tools that can be used. We believe this will help our users enhance an ever-growing set of workflows in their day-to-day.

Multi-tool assistants can handle more advanced workflows, no matter how many sub-steps might be needed.

Assistants previously used zero or one tool. Assistants previously operated with either no tool (the assistant would just rely on the model and assistant instructions to generate a response) or one tool (the assistant could use a single configured tool and then generate a response based on the output of that tool).

Assistants can now be configured to pick and choose from a set of tools and use several of them at once. This could, for example, allow you to automatically draft a sales follow-up email by (1) finding some news about your prospect on the Internet, then (2) extracting the key highlights of your latest calls from your internal notes, and (3) draft an email with a specific tone and format, using content from (1) and (2).

Dust lets you define the tools available to an assistant and inspect how they were run.

  • When called, the assistant analyzes your request and decides which configured tools will best serve your query. It then runs a flexible combination of tools, from none to several, at each step of its reasoning process.
  • The tools used by the assistant are available for you to inspect, to have visibility on how the response was produced.

What are all the tools assistants can use?

Assistants are now able to pick from:

  • Semantic Search - this one surfaces the pieces of your internal data that are relevant to answer to your question without opening any document
  • Web Navigation - use this to incorporate live information from the internet in your answers
  • Data analysis with Query Tables - allows you to navigate and make computations from your spreadsheets without code or formula, just conversation.
  • Unstructured data processing (with methods called Extract Data and Most Recent Data) - extract and analyse only the data you want from large sets of text, tickets, messages.
  • Any tool you design as a Dust app - You can also create your own actions with Dust apps, using our developer platform.
Assistant creation interface

On to some real-world applications

We’ve listed a few examples of tasks that require insights combining historical data, external real-time information, or structured data from internal data sources. Below are a few we identified and tackled.

Use Case Description of the task you would like to perform Multi-tool assistant configuration
[Sales / Product] Market research

You must check and report on your company's positioning to update your internal communications/strategy/sales pitch. To do this, you usually:

  1. Gather the most up-to-date internal information about your product and pricing rationale.
  2. Map and monitor your main competitors' websites and announcements.
  3. Add some of your internal key metrics that are stored in a large spreadsheet to enrich the benchmark.
  4. Create a sharable summary table comparing your key features and messages to your competitors with the links and numbers you gathered above.
  1. Semantic Search on your relevant Notion pages, Google documents, and Slack threads.
  2. Web Navigation on your competitor's websites.
  3. Query Tables on your spreadsheet.
[Knowledge Management] Surface the outdated parts of your knowledge base

You are part of a team where processes are documented on Notion, and you are in charge of keeping this content up to date. Every now and then, new team members ask questions on Slack about complex, missing, or outdated content.

  1. You extract relevant data from Slack conversations.
  2. You scan through the different documentation pages on Notion to identify the relevant parts to update.
  3. You go ahead and write the edits, and maybe share what you changed with the team in a separate message.
  1. Unstructured data processing with Extract Data, scanning through Slack, and surfacing relevant messages only.
  2. Semantic Search on your Notion pages to identify the relevant parts to update.
[Engineering] Get coding guidance informed by your codebase and internal and external documentation

You are working on a new code project. At times, you need to check your company's internal documentation to move forward; some other times, you need advice from the public documentation of the tools and language you are using.

At the end, you also need to assess the impact of your project on your company's broader code base to communicate and roll it out properly. This all takes some time.
  1. Semantic Search on your internal documentation.
  2. Web Navigation on public documentation or forums.
  3. Semantic Search through your GitHub code base to identify related parts.
[Product / Operations] Project management

Like every week, you must send a project update to your team and communicate about your advances internally. To do so, you need to:

  1. Gather team members' achievements/blockers/plans on Slack.
  2. Check and report on tasks' status in a Notion database or a Gsheet.
  3. Summarise the notes from the latest team meetings.
  4. Add the latest update of the key metrics you track.

You know where to look, but it does take time to copy-paste, summarise, and format everything.
  1. Extract Data to generate a summary with all the right messages from your team members on Slack.
  2. Extract Data to list status labels and summaries of content from a Notion database.
  3. Semantic Search through your transcripts on Gdrive to summarise notes.
  4. Data analysis with Query Tables to add the most up-to-date figure from your tracking Gsheet or CSV.
[Product / Operations] Unstructured data analysis

You have a large set of customer comments to analyze: say you want to perform an analysis to understand why some users churned. To do so, you would like to:

  1. Map each customer to a certain category, taken from your internal customer segmentation documents.
  2. Get a list of tickets from a specific type of customer and content.
  3. Generate a short summary with the top 3 reasons for churning with relevant customer quotes as highlights at the end.
  1. Semantic Search to map each customer to your internal categories.
  2. Extract Data to extract only relevant tickets to summarise at the end.
[Sales] Drafting outbound / follow-up emails

As we suggested above, you want to find some news about your prospect on the internet, then extract a summary of your last calls from your note and have your email written with the specific tone and format you like.
  1. Web Navigation: the specific information you want to get about your prospect.
  2. Semantic Search to get the highlights from your notes in Gdrive or Slack.

More tools with Dust apps

With Dust apps, you can give assistants access to a longer set of capabilities today. Here are some examples:

We can’t wait to see the new use cases you will uncover!