Skip to main content

Workflow

You can access your workflow by clicking on Agent workflow in the left navigation menu.

It is structured into 4 main steps:

  • Preprocess: In this step, you can rephrase the user question or enrich it in order to make it more efficient and more likely to produce relevant matches when performing the retrival step.
  • Retrieval context: This step will run all the information retrieval needed to produce the answer.
  • Generation: Here, you use the retrieved information to produce the expected output.
  • Postprocess: In this last step, you can assess the quality of the output or trigger external actions.

Preprocess​

In Preprocess, Rephrase is only available agent at the moment.

It rephrases the user question in order to perform better in the retrieval step. You can orient the way the rephrasing is done by specifying some rules. You may also specify a Knowledge Box that will be searched to collect relevant information to rephrase the question.

Retrieval context​

Ask​

This agent is able to perform a call to your Knowledge Boxes defined in your drivers.

You can select one or several Knowledge Boxes, and then you need to enter the search and RAG parameters to apply.

Perplexity and Google​

These two agents let you call the Perplexity and Google defined in your drivers.

MCP​

This agent makes MCP calls to the MCP servers defined in your drivers.

SQL​

This agent uses the SQL driver to perform natural language queries over the configured relational database.

This agent is currently in beta.

From the UI, you can configure several options for this agent:

  • Conversion model: Select the LLM used to translate the user query into a SQL statement.
  • Dynamic table selection: Allow the LLM to pre-filter and route the relevant tables automatically, saving context tokens. You can also specify a custom Table selection model for this logic.
  • Tables filtering: Narrow down the data the agent has access to by specifying include_tables or ignore_tables.
  • Context Details: Configure how much structure to pass to the LLM via options like injecting sample rows (sample_rows_in_table_info) or providing index information (indexes_in_table_info).
  • Retries: Adjust the number of attempts the agent will make if the generated SQL syntax is invalid.

Snowflake​

This agent uses the Snowflake driver to perform natural language queries over the configured Snowflake database.

This agent is currently in beta.

It works similarly to the SQL agent, but applies Snowflake-specific SQL dialect guidance and cost controls. The agent uses dynamic table selection by default, so it can choose the most relevant tables before generating and running a query.

From the UI, you can configure several options for this agent:

  • Data source: Select the Snowflake driver to query.
  • Conversion model: Select the LLM used to translate the user query into a Snowflake SQL statement.
  • Tables filtering: Narrow down the data the agent has access to by specifying include_tables or ignore_tables.
  • Schema: Optionally override the default Snowflake schema configured in the driver.
  • Max result rows: Limit the number of rows returned by each Snowflake query. The agent also injects a LIMIT clause when the generated SQL does not include one.
  • Retries: Adjust the number of attempts the agent will make if the generated SQL syntax is invalid.

Pandas​

This agent leverages a pandas dataframe (e.g. from an uploaded .csv data) to answer queries.

This agent is currently in beta.

From the UI, you can configure various settings:

  • Data Source: Input spreadsheet data directly (in .csv format).
  • Conversion model: Select the LLM that will translate the natural language query into dataframe manipulation operations.
  • Context Details: Provide the LLM with an understanding of the structure by specifying the number of sample rows (sample_rows_in_table_info) or adding index details.
  • Retries: Number of retry attempts in case the generated pandas code is invalid or raises errors.

Generation​

In Generation, Summarize is only available agent at the moment.

It takes all the retrieved content from the previous step and phrases an answer according the user question.

Postprocess​

Validation​

This agent uses the REMi model to estimate the validity of the answer.

External Call​

This agent makes an HTTP call to any external endpoint, that way the result of the workflow can be used to trigger an action.