Knowledge tools

Status: Current | Last reviewed: 2026-09-10

A Knowledge tool turns a content source into a searchable tool for agents and workflows. TrueServe creates a Knowledge Source, indexes the content, and exposes one callable operation: search.

Knowledge tools appear only when your tenant admin has enabled and configured Knowledge under Settings → Tenant Settings → Knowledge. If you do not see Knowledge in Add tool, ask your tenant admin to turn it on and supply the Knowledge Sources address and API token — see Knowledge settings.

Create a Knowledge tool

  1. Open Tools & Templates and select the Tools tab.
  2. Select Add tool, then choose Knowledge.
  3. Name the tool and choose a source type.
  4. Fill in the source details and authentication fields.
  5. Configure the visible Indexing and retrieval defaults. Expand the Advanced sections only when the source needs specialized parsing, filtering, GraphRAG, or RAPTOR behavior.
  6. Save the tool. TrueServe creates the source and starts indexing when the source supports it.

The source identifier is handled automatically from the tool name — there is nothing to fill in.

After saving, the form shows source status, sync state, indexed document count when available, and the last error. Use Sync now to reindex a saved source.

Create an inline Knowledge tool

Agents and workflow Tool nodes can also create a Knowledge tool directly from the tool picker. Select Create Inline Knowledge Tool, then choose the source type and source details in the same form used by shared Knowledge tools.

Inline Knowledge tools are saved only on that agent stage or workflow. They do not appear in the shared Tools section, but they still create a Knowledge Source, support status polling and file upload management, and use your tenant's Knowledge API token at runtime. Neither shared nor inline Knowledge tools store a copy of the address or token — they always use whatever your tenant admin has configured, so updating the settings takes effect everywhere immediately.

Source types

Source What you provide
Website One or more URLs, crawl options, and an optional crawl frequency: Hourly, Daily, Weekly, or Monthly. See Website crawl options.
Files A source prefix, uploaded files, and whether to reindex after upload. Saved file sources show a file list with delete actions.
Amazon S3 An AWS integration, bucket name, region from the integration, and optional prefix.
Google Cloud Storage A Google integration with GCS-compatible HMAC access keys, bucket name, and optional prefix.
Azure Blob Storage An Azure integration with storage account credentials, container name, and optional prefix.
Confluence Confluence base URL, username, access token, cloud/server mode, and the space or page to index.

For S3, GCS, and Azure, credentials come from the selected integration. For Confluence, the access token is pass-through only: TrueServe sends it to the Knowledge API during setup or credential-changing updates, but does not store it in the tool config. You must enter it again for later updates that need Confluence credentials.

Indexing and retrieval

The creation form loads its choices and defaults from the Knowledge Sources API. Model lists therefore show only embedding, reranking, generation, vision, and OCR models configured for the current RAGFlow tenant. If the live catalog is temporarily unavailable, the form keeps documented non-model defaults available and preserves saved model IDs.

The main controls are:

Setting What it controls
Document type Parser strategy for general documents, Q&A, tables, books, laws, presentations, pictures, and other specialized content.
Embedding model Vector model used for indexing and queries. Use tenant default leaves model selection to RAGFlow. A selected model may show a recommended chunk size and input limit.
Chunk size Target tokens per indexed passage. Start with the model recommendation or 512.
PDF parser DeepDOC or a configured vision/OCR model used to extract PDF content.
Default results Number of passages returned when the caller does not choose a count.
Similarity threshold Minimum relevance. Raise it for precision; lower it for recall.

Ingestion settings affect parsing and indexing. Existing documents generally need a new sync or reparse after a change. RAGFlow normally prevents changing the embedding model after chunks exist because one dataset cannot mix vector dimensions.

Advanced Knowledge settings

Advanced settings are grouped so ordinary sources can keep safe defaults:

GraphRAG and RAPTOR are model-intensive secondary indexes. Enabling them stores their configuration; build them after normal source parsing completes. If both are enabled, they use the same generation model.

The Agent can… switches are policy controls, not model selectors. For example, Agent can toggle reranking exposes a boolean choice but never exposes the approved reranking model ID. Graph, TOC, and metadata filtering follow the same rule.

Website crawl options

A Website source can index one site or several. Enter a starting URL, then choose Add another site to combine multiple sites or sections into a single knowledge source — all sites are crawled into the same index and searched together.

For each site:

Option What it does
Website URL The starting URL to scrape or crawl.
Crawl linked pages Off indexes only the single starting page. On follows links and reveals the crawl options below.
Max pages Caps how many pages the crawl indexes. Leave blank for the default.
Max link depth How many link hops from the starting URL to follow. Leave blank to let the crawler decide. Use 2–3 for gateway pages whose links lead into rich sub-trees.
Crawl the entire domain On follows links across the whole domain. Off keeps the crawl under the starting URL's path — use it to scope a crawl to one subdirectory.
Include paths One anchored path pattern per line (for example ^/docs). Only pages whose path matches are kept.
Exclude paths One anchored path pattern per line (for example ^/blog). Matching pages are dropped, applied after include paths.

Include and exclude paths are anchored regular expressions matched against the URL path, so ^/docs matches /docs and everything under it. Leave both blank to keep every crawled page.

Uploaded files

Choose Files when you want to upload documents directly. You can select multiple files at once. TrueServe sends each file to the Knowledge API and, by default, triggers a reindex after upload.

Saved file sources show their current file list. Delete a file from this list when it should be removed from the indexed source, then sync the source if needed.

Use the tool

Attach the Knowledge tool from the normal agent or workflow tool picker. Shared Knowledge tools expose a search function; inline Knowledge tools expose the tool name you entered. The base tool always requires query. Additional parameters appear only when their Agent can… switch is enabled:

Parameter Notes
query Required search text.
top_k Result count, 1-50. Exposed by default.
similarity_threshold Relevance threshold, 0-1. Hidden from the agent by default.
vector_similarity_weight Semantic-versus-keyword balance, 0-1. Expert control; hidden by default.
Reranking, graph, TOC, and metadata controls Exposed only when the corresponding feature is configured and its bindable switch is on.

Use clear tool names such as "Search support articles" or "Search policy documents" so the model knows when to call the tool.

Troubleshooting