What the data for this category looks like
Sources primarily include internally archived past official speeches and special report materials from central state-owned enterprises and public sector entities, plus official document standards and policy interpretation documents issued by superior authorities. Update frequency is adjusted according to business cycles. Batch additions of current reference materials are made before quarterly and annual reporting periods, and corresponding interpretation documents are added synchronously when policies are updated. Most documents follow a structure including topic, main body, and closing information. Some include additional fields such as meeting attendees and report recipients. Single-document length varies widely: some special report materials can reach several thousand characters, while most regular speeches are several hundred characters long.
Constraints on Knowledge Base Retrieval and Recall From These Characteristics
Internally archived speeches and report materials have inconsistent formatting. Some older documents lack standardized typesetting, which interferes with the accuracy of chunking and semantic recognition. Update frequency is adjusted according to business cycles, so incremental synchronization on demand must be supported to avoid redundant data from full retrieval. Documents include additional fields such as topic and report recipient. Retrieval must first match topic keywords, while associating context information to ensure semantic completeness. Single-document length varies widely. Long documents must avoid overly fine chunking that splits information, while short documents must avoid overly coarse chunking that loses key details.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches topic relevance for speeches and report materials, avoids introducing irrelevant results if set too low, and avoids missing relevant historical materials if set too high |
limit | top 8–12 results | Reference materials for a single drafting task should not be excessive, to avoid context overflow while covering main reference directions |
searchMode | embedding | Prioritizes matching semantically related historical documents, adapts to the topic consistency requirement for speeches, replacing the rigid results of keyword matching |
usingReRank | Enabled | Performs secondary ranking on recall results, corrects deviations in semantic matching, and ensures highly relevant materials are returned first |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands keywords from the user's drafting request, such as expanding "quarterly report" to related sub-scenarios, to match more accurate historical documents |
collectionIds | Specify collection IDs for the exclusive official document knowledge base | Limits retrieval scope to the exclusive official document knowledge base, avoiding mixing in irrelevant business data |
Three Common Mistakes
- Retrieval results do not return image content from Word documents. Cause: Default retrieval only extracts text content, and no image association display or OCR parsing logic is configured, so image information cannot be properly recalled or returned.
- After upgrading to FastGPT v4.17.0, an error
cannot fetch internal urlis returned when creating a knowledge base using a URL. Cause: The new version restricts access to internal URLs, and no allowed internal domain whitelist is configured, so the target URL content cannot be crawled. - When deploying the open-source version locally, it is not possible to create more knowledge bases than the default limit. Cause: The open-source version has a default knowledge base quantity limit, and the restriction parameter has not been modified via the corresponding configuration file.
How to Verify Proper Configuration
- Submit a test query for a drafting request, check whether the number of returned retrieval results matches the configured
limitrange. - View the relevance ranking of retrieval results, confirm that highly relevant materials are displayed first, and verify that
usingReRankis active. - Upload a Word document containing images to the test knowledge base, initiate a corresponding query, and confirm that image content is properly associated or displayed.
- After configuring
collectionIds, initiate a test query, confirm that retrieval results only come from the specified official document knowledge base collection, and no other business data is included.
The values given are common starting points and should be measured against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.