Contextual AI n8n integrationAutomate Contextual AI with n8n.

One run sets up an agent, its datastore and the first documents. The response carries the agent identifier you reuse in every later operation.

• Agent Name: the name the new agent gets, for example a customer support bot.
• Configuration Mode: Pre-Built uses an Agent Composer template, Custom takes your own YAML in Custom Template.
• Pre-Built Template: Simple Search for fast multi-modal retrieval, Agentic Search for multi-step reasoning and tool use.
• Datastore Configuration: create a store with Datastore Name, or attach existing ones with Datastore IDs.
• System Prompt: the instructions the agent references when it generates a response.

Reads back every agent on the account with its identifier, so a workflow can look one up by name instead of carrying a hard coded ID.

• Limit: required, the maximum number of agents returned in one call.
• Cursor: the value returned by the previous call, used to fetch the next set of results.

Removes an agent for good. The operation is irreversible, so it belongs behind a condition rather than in the middle of a happy path.

• Agent ID: required, the identifier of the agent to delete, usually {{ $json.id }} from a List Agents step.

Sets up an empty store that documents get ingested into later. Useful when the datastore has to exist before anything else runs.

• Datastore Name: required, the name of the store, for example the one holding product documentation.
• Configuration (JSON String): the store configuration as stringified JSON; left empty, the default configuration applies.

Pushes a file from the incoming item into a datastore, where it becomes searchable material for an agent.

• Datastore ID: required, the store the document goes into.
• Input Binary Field(s): the binary field name on the incoming item; left empty, every binary field is ingested.
• Metadata (JSON String): a stringified JSON metadata request whose custom_metadata is a flat set of key and value pairs.

Reads the metadata and the processing status of one document, which is how a workflow finds out whether ingestion actually finished.

• Datastore ID: required, the store that holds the document.
• Document ID: required, the identifier of the document being inspected.

Brings back the stores on the account, optionally narrowed to the ones an agent is linked to.

• Limit: required, the maximum number of stores returned in one call.
• Cursor: the value from the previous call, used to walk to the next page.
• Agent ID: filters the list down to the stores linked to that agent.

Sends a question to an agent and returns its answer, grounded in the documents that agent can see.

• Agent ID: required, which agent answers.
• Query: required, the question itself, often {{ $json.question }} from a form.
• Retrievals Only: on, the operation fetches retrievals and skips generation.
• Include Retrieval Content Text: adds the text of the retrieved contents to the response.
• Conversation ID: continues an existing conversation; set it and the conversation history is ignored.

Scores a set of texts against a question and returns them in order of relevance, so only the useful ones move on.

• Query: required, the question the documents are scored against.
• Documents: required, a comma separated list of document texts.
• Model: ctxl-rerank-v2-instruct-multilingual, its mini variant, or ctxl-rerank-v1-instruct.
• Top N: how many top ranked results come back; leave it at 0 to return all of them.
• Instruction: an optional steer for the reranker.

Grades a response against a written rule, which turns a vague feeling about answer quality into something a workflow can branch on.

• Query: required, the question that was asked.
• Response: required, the answer being evaluated.
• Unit Test: required, the rule the response has to satisfy, written as a sentence.

Starts a parse job on a file carried by the incoming item. It returns the job, not the parsed text, so the workflow continues with the two operations below.

• Input Binary Field(s): the binary field name on the incoming item to parse.
• Parse Mode: Standard is the mode on offer.
• Figure Caption Mode: Concise or Detailed captions for figures.
• Enable Document Hierarchy: keeps the structure of the document.
• Page Range: narrows the job to a slice of pages.

Tells you where a parse job stands. It is the polling step between starting a job and collecting what it produced.

• Job ID: required, the identifier of the parse job to check, taken from the operation that started it.

Collects the output of a finished parse job and hands it to the rest of the workflow.

• Job ID: required, the identifier of the parse job to retrieve results for.
• Output Types: a comma separated list among markdown-document, markdown-per-page and blocks-per-page.