What the data for this category looks like
This category's data is primarily sourced from the national government service platform's "One Thing Done Efficiently" special zone, local integrated government service platforms, electronic certificate repositories, and various departmental approval systems. The update frequency is daily. Commonly involved documents include joint handling item lists, one-time notice sheets, joint handling application forms, material circulation slips, parallel approval opinions, completion receipts, and more. These documents contain cross-departmental item association fields, material reuse requirements, handling circulation nodes, and other relevant content. Some documents are exported from structured systems, while others are unstructured text or scanned documents.
What constraints do these characteristics impose on the knowledge base retrieval and recall link?
Scattered data sources and daily updates require the retrieval link to support multi-dataset association and incremental synchronization configuration. Documents with cross-departmental association fields require retrieval to prioritize matching contextually relevant content for joint handling items. The diverse document types, including both structured fields and unstructured text, require retrieval modes to balance vector matching and keyword matching. Joint handling items involve multi-department circulation, so recall results must cover multiple document categories such as pre-approval and material requirements, to prevent process stalls caused by missing context. Daily updated data also requires the retrieval system to quickly respond to the latest item adjustment information, to avoid using expired documents.
How to set the configurations
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Joint handling documents have many association fields, need to filter low-relevance non-joint handling content |
limit | Top 8–12 entries | Joint handling items involve multi-department documents, avoid context overload that affects model understanding |
searchMode | Hybrid retrieval | Balance structured field matching and vector recall for unstructured text |
usingReRank | Enabled | Multi-source recall results need ranking optimization, prioritize returning high-value documents related to joint handling |
datasetSearchUsingExtensionQuery | Enabled | Joint handling questions have complex phrasing, extended query terms can cover more associated documents |
collectionIds | Bind dedicated joint handling dataset IDs | Avoid irrelevant government documents interfering with retrieval results for joint handling scenarios |
Three common mistakes
- Issue: Retrieval results fail to match joint handling item documents in the local knowledge base. Cause: Joint handling item lists, approval opinions and other dedicated documents are not included in the specified
collectionIdsrange, ordatasetSearchUsingExtensionQueryis not enabled, leading to query terms that do not adapt to the complex phrasing of joint handling scenarios. - Issue: The number of recall results returned in a single batch is far lower than the set
limitvalue. Cause: Some PDF documents have not been parsed, or thesimilaritythreshold is set too high, filtering valid recall content. - Issue: Scanned PDF documents cannot be retrieved. Cause: The documents have not undergone OCR processing, so no corresponding vector embeddings can be generated, making them unable to participate in retrieval matching.
How to confirm the configuration is correct
- Upload a joint handling item list document to the specified dataset, retrieve a query term that includes the item number, and verify that the returned results contain relevant content from this document.
- Enter the dataset configuration page of FastGPT v4.17.0, confirm that
collectionIdsis bound to the dedicated joint handling dataset, and does not include irrelevant government documents. - Test the recall result ranking before and after enabling
usingReRank, confirm that documents related to joint handling have a significantly improved ranking after reranking. - Upload a non-encrypted PDF joint handling application form, check the parsing log to confirm that the document has successfully generated vector embeddings.
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.