What the data for this category looks like
Data sources include scanned paper documents and structured forms generated during expressway operations, such as toll declaration forms, road administration files, toll ledgers, and overload detection records. Updates are conducted in batches aligned with business review cycles. Materials within a single cycle do not change frequently. Document structure includes fixed fields and attachments. Fixed fields include file number, traffic section, declared amount, detection time, and other standard entries. Attachments are mostly high-definition scanned paper materials. The text length of individual materials varies widely.
What constraints these characteristics bring to the knowledge base retrieval and recall link
The structured nature of fixed fields requires retrieval to balance exact field value matching and semantic matching of attachment content. Relying solely on semantic recall may lead to missing field information. The wide variation in individual material text length can result in redundant recall results or lost critical information. Batch-updated data requires incremental index update logic. This prevents mixing of old and new data, which would cause review errors. Road administration files and toll declaration forms have large differences in their fields. Distinguish retrieval weights for different document types to prevent interference between materials of different business types.
Configuration Settings
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Matches the mixed retrieval needs of structured fields and semantic content for expressway declaration materials, filters low-relevance results |
limit | Top 8 entries | A single file review material contains an average of 3 to 5 core fragments. 8 results cover all associated materials required for complete review |
searchMode | hybrid | Balances exact matching of structured fields and semantic recall of attachment text, adapts to business data with mixed structures |
usingReRank | Enabled | Semantic recall of long-text materials is prone to local matching deviations. Reranking improves the sorting priority of core information |
datasetSearchUsingExtensionQuery | Enabled | Performs extended retrieval on vague expressions in declaration forms, such as "special passage application", to cover synonyms and industry terminology |
collectionIds | Specified by business-specific collections | Limits the retrieval scope to only the knowledge base related to expressway operation declaration materials and file review, excludes unrelated documents |
Three Common Errors
- Phenomenon: The number of retrieval results is far lower than the preset
limitvalue. Logs show that the number of valid recalls is less than 3. Cause:datasetSearchUsingExtensionQueryis not enabled. This fails to cover industry terminology and vague expressions in declaration materials, leading to an overly narrow matching scope. - Phenomenon: A single query returns more than 20 redundant results. Core review fields cannot be located quickly. Cause: No reasonable
similaritythreshold is set, andusingReRankis not enabled. This causes low-relevance long-text fragments to be recalled first. - Phenomenon: After uploading a 10,000-line long-text file, no matching results are returned for queries. Cause: No reasonable chunking is performed for long texts, and
datasetSearchUsingExtensionQueryis not enabled. This causes the embedding vector to fail to cover the core content of the full text.
How to Confirm Proper Configuration
- Check the
collectionIdsconfiguration of the knowledge base collection. Confirm that only the dedicated collections related to expressway operation declaration materials and file review are included. - Submit a test query that includes a specific file number. Verify that the
similarityscores of returned results fall within the preset range. - Upload a 10,000-line test file. Submit a query and confirm that returned results include matching content for core review fields.
- After enabling
usingReRank, compare the result ranking before and after reranking. Confirm that the position of core information has been optimized.
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.