Knowledge Extraction Agent

The Knowledge Extraction Agent answers requests that need every matching line from many documents, not the few most similar passages a knowledge search returns. A user asks in one chat message, for example "Find all billing positions regarding hardware on the 2025 bills from Acme", and gets back one table: one row per record found, each row naming the path and file it came from, with a summary of how many files were read, how many rows were found and which files could not be read.

When to reach for this agent

Use it when the answer is a list of records collected across documents: invoice lines, contract parties, service levels, deadlines. For a question one passage answers, the Document Intelligence Assistant is faster and cites its sources. The Universal Agent offers the same extraction as one of its tools when its profile switches the document tools on.

What it does

1. Find the documents. The model works through the documents of the profile's collections with three tools: it lists their paths, searches their text for exact terms such as a supplier name or a year, and reads a document when it is unsure. It stays within the profile's collections and path filter at all times.
2. Select them. When it knows the documents, the model describes them, for example "invoices from 2025 that mention Acme", together with the columns the user asked for and the rows to keep. It sees how many documents that is and adjusts the description when the count or the examples are wrong.
3. Design the columns. The columns are designed from the user's words, such as date, item, amount, currency. A condition like "only hardware" decides which rows are kept and never becomes a column.
4. Extract. Every selected document is read and its matching records are extracted, many documents at the same time. Documents are read in overlapping parts, so a line on the last page is found and a line at the boundary between parts is not counted twice. A row that may not meet the condition is kept and marked unsure in a Match column, so a person can check it instead of it being dropped.
5. Answer. The table streams into the chat, followed by the summary. A document that could not be read is listed under the table with the reason; the other documents' rows are unaffected. Documents that were read but gave no rows are named under "No rows from", so an empty result never looks like a missing clause.

The chat shows each step as it happens: the tools the model calls while it looks for documents, a line saying how many documents will be read, which columns the table has and about how long it takes, and one line per document as it is extracted ("12 of 25 · invoice.pdf: 3 rows").

When more documents match than one answer extracts

Nothing is cut silently. When the documents would take longer than about five minutes to extract (Minutes per Answer), or are more than Files per Answer, the agent does not extract them all at once. The answer shows:

• how many documents match and how long extracting them all would take, about;
• how they break down by folder, by year in the path and by collection;
• a sample table from three of them, so you see the columns and rows before committing;
• what you can do: narrow the request, for example to one folder or year, or reply continue to extract all of them in parts.

Each continue extracts the next part and answers with its table, its place ("Part 2 of 4: documents 26–50 of 98") and how to go on; the last part gives the totals. The parts keep the columns of the first. Asking something new instead replaces the unfinished extraction. Above Most Matching Documents (1000 by default) only narrowing is offered; the answer then says "more than 1000" and lists the folders and years of the first 1000 documents by path without counts, since the rest were not looked at.

Continuing needs the same conversation: the unfinished extraction is kept with the chat, so a client of the OpenAI-compatible API passes the conversation's thread_id in metadata with every message, as the chat UI does.

The same holds for rows: when one answer finds more rows than Rows in the Table, the rest are shown first when you reply continue.

Long documents

A document longer than Longest Fully Read Document (500,000 characters by default, a few hundred pages) is read only in its relevant parts: its opening, which states what every row shares, and the parts where the wanted lines most likely are. Its rows are marked (partly read) next to the file name, and the summary lists it under "Partly read (rows may be missing)", because a line in a part that was not read is not found. See the ADR 2026_10_21_partial_reads_of_long_documents_in_record_extraction.

What it does not do

• It does not compare or reconcile documents. Matching an invoice line to its contract line, or flagging deviations, is a different task.
• It does not ask for approval before extracting. The extraction starts as soon as the documents are selected.
• It does not deliver the table as a file. The table is part of the answer; copy it or ask a follow-up question.
• It never reads outside its scope. Only the profile's collections are used, narrowed to those the asking user may read and to the profile's path filter. A request for other documents, or a # reference to a collection the profile does not list, is left out and the answer says so.

Limits

Every request is bounded, and nothing is cut without the answer saying so:

Limit	What happens at the limit
Maximum Decisions	The search for documents stops. If the agent had not chosen documents yet, it selects them by folder and checks the request's conditions in each; if it had, documents may be missing. The answer says which.
Maximum Tool Calls	As above.
Minutes per Answer	Larger selections get a sample and the offer to narrow or to continue in parts.
Files per Answer	As above, for the number of documents.
Most Matching Documents	Above it, only narrowing is offered.
Rows in the Table	The other rows follow when you reply continue.
Longest Fully Read Document	A longer document is read in its relevant parts and its rows are marked partly read.
Model Calls at Once	How many extraction calls one agent instance makes at the same time; bounds the load.

One fixed limit applies as well: the rows of one document travel in one message, so a document with very many long rows (around 40 rows of full contract clauses) keeps the first ones, and the summary names the document and how many rows were left out. If a collection the extraction read from is no longer available when you continue, the extraction ends with a notice; ask again to extract from what is available.

Setting it up

1. Grant the blueprint. The Knowledge Extraction Agent is not part of a new tenant's default set; a sysadmin grants it (see Access control).
2. Create a profile. In Admin > Agents > Blueprints, select Knowledge Extraction Agent and click Create Profile, or start from the Document Tables template.
3. Choose the collections. Under Knowledge Tool, list the databases and collections the agent works on. Saving checks that you may read them.
4. Set a path filter (optional). Under Knowledge Documents, a pattern such as invoices/** limits the agent to matching documents; it can never look outside it.
5. Choose the models. The main model looks for the documents and designs the columns; pick one that is reliable at tool calls. The Extraction Model reads each document; a smaller, cheaper model works well.
6. Review the limits and save.

Configuration reference

Field	Default	Description
Instructions	(empty)	Added after the built-in instructions, e.g. how documents are named.
Knowledge Tool > Collections	—	The databases and collections the agent works on.
Knowledge Documents > Document Path Filter	(empty)	Only documents whose path matches this pattern, e.g. invoices/2025/**.
Table Extraction > Extraction Model	main model	The model that reads each document and extracts its rows.
Table Extraction > Files per Answer	100	The most documents one answer extracts from.
Table Extraction > Minutes per Answer	5	About how long one answer may extract before the rest is offered.
Table Extraction > Most Matching Documents	1000	Above it, the user is only asked to narrow the request.
Table Extraction > Rows in the Table	500	The most rows one answer shows; the rest follow on continue.
Table Extraction > Longest Fully Read Document	500000	Longer documents (in characters) are read in their relevant parts.
Table Extraction > Model Calls at Once	4	Extraction calls one agent instance makes at the same time.
Tools > Maximum Decisions / Maximum Tool Calls	10 / 20	How long the model may look for documents.