Knowledge Management Agents¶

Last reviewed: 2026-10-09

Knowledge Management Agents prepare documents for search, test answers against source evidence, and publish a selected knowledge release. The feature must be enabled by your administrator. Its initial destination is native RAGFlow.

Create and prepare knowledge¶

Choose Knowledge Management Agent when creating an agent. In Prepare, add website addresses, upload documents, or select a folder. Supported inputs include HTML, Markdown, text, PDF, Word, Excel and images. The file limit is 100 MB. Edit a source to set its website scope and typed metadata. Use effective_from and effective_through for known applicability dates; leaving a date unknown is different from declaring it applicable to every date.

The canvas starts with Sources → Parse → Split → Validate → Native RAGFlow index. Select a step to change its controls. Add AI enrichment or transformation when needed, use Decision to route by document/metadata fields, and join branches with Merge. Every branch needs validation after its final change. Undo/Redo and the list view provide alternatives when arranging the canvas.

Select a preparation/answer model in Answer. Scanned pages and images also need a vision model in Parse. A configuration administrator can set the native RAGFlow connection in Health. Save, then choose Build knowledge. You can leave and return while the worker runs. Failed builds show the cause and offer Retry; completed preparation is retained. Cancel stops subsequent work. Graph retries also reuse saved document extraction and already-merged work. Graph jobs show their current phase: extraction, reusing saved evidence, repairing records, merging, entity resolution or community summaries. These phases can take longer than indexing; the verified evidence count is not a graph-completion percentage. After an interrupted write, Repairing graph records completes the saved work before merging continues. Cancellation lets an active database write finish safely. Text uploads must use UTF-8, or UTF-16 with a byte-order mark. Binary files sent as text are rejected; recognized images require an OCR model.

Try answers and review evidence¶

Select a completed build, open Answer, and ask a question. The selected build uses its saved settings. Build again to test configuration edits.

Controls include query rewriting, retrieval depth, similarity weighting, reranking, applicability filters, reviewed terminology and graph expansion. Graph expansion requires a completed native graph step. Answers show cited source evidence and missing information. Expand evidence to inspect its original text and source page, sheet or row references. Missing information should identify what prevents answering your question. Optional next steps or adjustments you explicitly excluded should not appear as gaps. Check that citations support both the conclusion and any added detail. Initial evidence units controls the first retrieval. Graph links may add more evidence, up to 50 units in total. Large graph candidate sets are ranked within your selected filters; the answer shows how many candidates were found and added. Use narrower filters when a question needs more focused evidence.

Health shows document/unit counts and native graph status. Inspect prepared evidence to compare source text with the searchable representation. A successful build proves that its configured processing completed; it does not prove answer quality.

Inspect graph shows matching entities and their connected relations. Expand an entity or relation, then Original source support, to compare its generated description with source excerpts and page/row locations. Each view contains up to 50 entities, 100 relations and 100 excerpts. Excerpts are limited to 2,000 characters; narrow the search or inspect prepared evidence for more context. An assertion without visible source support still needs review.

Test knowledge¶

In Test, add a question and optional reference answer. You can generate candidate Q&A from an explicit source sample, paste a transcript, or select a prepared Analytics release. Transcript text is redacted before processing. Discovery includes follow-up questions and late turns; spoken answers do not establish the correct policy answer.

Review each candidate before marking it reviewed. Cases without reviewed requirements remain unmeasured. Choose Run evidence tests, then inspect individual results for the answer, explanation, missed claims and required-source recall. Compare results from different builds before choosing a release.

In Health, discover terminology candidates, inspect their supporting quotes, and approve or reject each mapping. Reviewed mappings enter the next build. Avoid approving similar words as equivalents when their policy scope differs.

Publish and refresh¶

Choose Publish tested release in Health after building the saved configuration and running its tests. Publication normally requires passing reviewed tests. The button becomes available when the selected build matches the saved draft and its required reviewed tests pass. A hint explains which step is missing. The workspace shows published generations; Restore selects an earlier eligible release and records a new publication.

Source edits or removal invalidate dependent releases until rebuilt. Restore cannot recover revoked source access. Enable scheduled refreshes and save an interval to create new candidates automatically. Publication remains a review step. A native graph export that is empty, failed or truncated must be fixed before a graph-dependent release can complete. Entity resolution must also finish every candidate decision. A failed or incomplete model response keeps the build from completing; Retry reuses successfully saved work.

Cloud-storage/Salesforce connectors, other index destinations and full audio/video preparation remain planned extensions. The source controls show the currently available inputs.

Review or import evidence requirements¶

In Test, select Edit requirement to correct a candidate question or answer, add required claims and source evidence, set an applicability date or typed metadata filters, and mark an expected insufficient-evidence result. Mark a case Reviewed against source evidence only after checking its references. Discovery context shows the supporting quotation and transcript window. A spoken answer does not establish a reference answer.

Import questions or Q&A accepts CSV with question and optional answer columns, including quoted commas and newlines. JSON accepts an array of case objects. Imports are limited to 3 MB in the browser and 500 cases per request; they always enter as candidates. Question-only cases remain unmeasured until you establish a reference answer, required claims or an insufficient-evidence expectation.

In a Decision step, choose document scope or route each split evidence unit. Unit routing can use extracted typed metadata from an earlier enrichment step.

Compare and consume a release¶

In Health, Compare source changes compares the selected candidate with an earlier publication or completed build. It shows changed sources and original files, added/removed evidence units, and cases or terminology mappings that refer to changed sources. Review those dependencies before promotion.

After publication, Copy MCP endpoint gives an endpoint for an ordinary MCP tool. Configure a tenant API token with agents:run scope in the consuming tool. ask_knowledge uses this agent's current publication and returns the same source evidence as the Answer tab. Dataset IDs and native credentials are resolved on the server.

Inspect and connect preparation steps¶

Select a step and choose Inspect this step, select a build document, then Load saved output. Parsed blocks and prepared units remain separate from original source evidence. The panel shows the selected build's saved artifact; new edits take effect in the next build. Native index and graph inspection is in Health.

List view provides keyboard-accessible source, target and branch selectors under Connect steps. Use Disconnect on an existing connection to rewire it. Decision outputs include Otherwise; all branches must rejoin explicitly and reach validation before indexing. The editor explains invalid connections before a build is queued.

The source and step panels scroll independently. Changing the selected build clears the previous build's answer, graph, impact report and inspected results.