What the data for this category looks like
Policy and regulatory basis data for power generation groups mainly comes from internal official red-headed documents, normative documents released by the National Energy Administration and industry regulatory agencies, and internal group management manuals. Data updates proceed in line with regulatory policy adjustments and group system revision rhythms, with no fixed cycle. Updates are synced immediately after revisions are completed. Most documents are in PDF or DOCX format. Their structures include fields such as document number, issuing entity, effective date, scope of application, and clause entries. Clauses use article, paragraph, and subparagraph as hierarchical units. The length of single documents varies widely. Some annual summary files can reach hundreds of pages.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The multi-source and non-standard format characteristics of power generation group policy and regulatory data require the retrieval link to support cross-format parsing and field-level precise matching, to avoid losing key information such as document numbers and effective dates due to format differences. The wide range of document lengths and some lengthy files will cause context breaks after embedding single data entries, so long document chunking rules need to be adapted. The lack of fixed update cycles requires the knowledge base to support incremental synchronous updates, to prevent retrieval results from lagging behind the latest policies. The dense use of specific industry terms also requires the recall link to prioritize content associated with professional terms.
How to set the configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Power generation policies have a high density of professional terms. This range filters low-relevance results and retains core relevant content |
limit | Top 8–12 results | Single policy documents are lengthy. Controlling the number of recalled results avoids context overflow, while covering scenarios with multiple associated clauses |
searchMode | Vector retrieval + keyword hybrid | Policy documents contain structured information such as clear document numbers and clause numbers. The hybrid mode can match both semantic and structured fields |
usingReRank | Enabled | There are many professional terms. Reranking further filters recall results with weak semantic relevance, improving answer accuracy |
datasetSearchUsingExtensionQuery | Enabled | Policy terms have synonymous expressions (such as "grid connection" and "grid access"). Extended queries can cover more relevant content |
collectionIds | Specify the exclusive collection ID for policies and regulations | Distinguish business data from other documents, avoiding interference from irrelevant content in retrieval results |
Three common configuration errors
- Phenomenon: Retrieval results do not include embedded images in DOCX-format policy documents, and supporting diagrams cannot be displayed. Cause: The knowledge base image parsing function is not enabled. Only document text content is extracted, and embedded charts, screenshots and other resources are not processed.
- Phenomenon: The number of results returned by a single retrieval far exceeds expectations, leading to redundant answers and context overflow. Cause: The
limitparameter is not set reasonably, with an overly large value range, recalling too many low-relevance policy clauses. - Phenomenon: Retrieval results do not match the latest revised policy documents. Cause: Incremental update trigger rules are not configured, and the knowledge base has not synced the latest system revision versions.
How to confirm the configuration is correctly applied
- Upload a DOCX-format policy file containing embedded images to the exclusive policy collection, initiate a test query, and check whether the returned results include content parsed from images.
- View the retrieval log to confirm that the number of recalled results matches the setting range of the
limitparameter, with no excessive redundant content. - Compare the newly released policy documents with the retrieval results, confirming that the returned content includes the latest revised clauses and effective dates.
- Check the collection configuration, confirming that only the specified
collectionIdsare included in the retrieval scope, with no irrelevant data from cross-collection mixing.
The values provided 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.