What this category's data looks like
Data mainly comes from official meeting communiqués, recorded leader speeches, and authoritative interpretation documents released by central and local authorities. Update schedule adjusts based on important meeting timelines. Core materials are added to the knowledge base within 1 to 3 working days after a meeting concludes. Daily learning support materials are updated quarterly. Most documents have long text structures, with fields including meeting theme, core points, deployment requirements, and interpretation notes. Single document length ranges from 1000 to 5000 characters. Some interpretation materials split paragraphs to mark core viewpoints.
What constraints these characteristics impose on knowledge base retrieval and recall
Long texts account for a large share. When chunking, retain contextual connections of core viewpoints to avoid losing the logical chain of meeting deployments after splitting. Core materials are updated concentratedly after important meetings. Retrieval must prioritize recalling documents added within the last 3 months to ensure alignment with the latest meeting spirit. Documents include structured fields such as theme and core points. Field restrictions can narrow retrieval scope, but avoid overly strict filtering that causes omission of authoritative interpretation content. Official expressions are rigorous and standardized. Retrieval must match standard usage of official keywords to prevent irrelevant content from being recalled due to semantic deviation.
Configuration settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meeting spirit content is rigorous. Requires matching official expressions with relatively high similarity to avoid recalling irrelevant content |
limit | Top 6–8 results | Single meeting material has lengthy content. Too many recalled results will exceed the context window. Too few will fail to cover all core points |
searchMode | Semantic + keyword hybrid retrieval | Balances precise matching of official keywords and semantic association, adapting to the rigorous expression and logical connections of meeting spirit |
usingReRank | Enabled | Semantic dispersion may occur after long text chunking. Re-ranking can prioritize chunks that best fit the meeting theme |
datasetSearchUsingExtensionQuery | Enabled | Can supplement official interpretation keywords related to the meeting, improving recall comprehensiveness |
collectionIds | Specify the exclusive collection ID for party building meeting materials | Isolates other irrelevant knowledge base content to avoid recalling unrelated documents |
Three common misconfigurations
- Phenomenon: Insufficient retrieval results, only 2 or fewer items are returned. Cause: Set an overly strict
similaritythreshold, or restricted the collection scope with filtering conditions, resulting in failure to recall authoritative meeting materials. - Phenomenon: Response timeout, with processing time exceeding 15 seconds. Cause: Did not enable
usingReRankand recalled too many items, large number of long text chunks leading to excessive re-ranking calculation load, or deployment resources not adapted to long text retrieval requirements. - Phenomenon: Retrieval results include non-official source meeting interpretation content. Cause: Did not specify
collectionIds, resulting in recalling documents from other non-exclusive collections, or enablingdatasetSearchUsingExtensionQueryintroduced non-official associated content.
How to confirm the configuration is correctly set
- Check the knowledge base collection's import records to confirm core meeting materials have been imported according to the update schedule.
- Initiate a test retrieval, enter meeting theme keywords, and verify that returned results originate from the specified collection.
- Adjust the
similaritythreshold, observe the matching accuracy of returned results, and confirm it meets the expected matching range. - Test the retrieval response time, and confirm it meets the business-set time threshold.
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.