What This Category of Data Looks Like
The data for refined oil sales declarations and file review comes from retail approval declaration forms, monthly sales filing files, inventory stock-in/stock-out ledgers, and compliance inspection archives. There are three update frequency categories: Daily declaration materials update in real time as approval workflow nodes move. Monthly filing files update once at the end of each month. Inventory ledgers sync daily. Most documents are in Word or PDF format with fixed fields, including unified social credit codes of business entities, oil product grades, actual sales volume, approval document numbers, rectification records and other fields. Units involve kiloliters, ten thousand yuan, batch numbers, business license numbers and other identifiers.
What Constraints Do These Characteristics Impose on the "Knowledge Base Retrieval and Recall" Link
Fixed fields and multiple update rhythms mean retrieval must prioritize matching precise fields such as business entities, oil product grades, and approval document numbers to avoid generalized recall of irrelevant data. Corresponding incremental synchronization rules must be configured for data sources with different update frequencies. This ensures recalled monthly filing files use the valid current month version, and inventory ledger data reflects the latest daily status. Declaration data for multiple oil product grades must be isolated by oil product grade dimension. This prevents mixed recall of declaration materials for different refined oil grades. Most file documents include compliance rectification records. Retrieval must associate workflow node keywords to ensure recalled review results match the current declaration stage.
How to Configure the Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Refined oil declaration materials are mostly structured compliance documents. This range filters low-match irrelevant files, retains accurately matched approval and filing content, and avoids filtering valid review records. |
limit | Top 8 entries | A single review needs to cover three types of data: associated approvals, inventory, and filings for declaration materials. 8 entries balances recall volume and context length, and complies with the context loading rules of FastGPT v4.17.0. |
searchMode | vector + fulltext | Retrieval must match both semantic vectors and keywords of structured fields. This covers precise retrieval needs such as oil product grades and approval document numbers, and adapts to retrieval scenarios for multi-field structured documents. |
usingReRank | Enabled | Most declaration files have similar compliance descriptions. Reranking places results most relevant to the current declaration topic first, improving review staff’s access efficiency. |
datasetSearchUsingExtensionQuery | Enabled | The system automatically expands oil product grades and business entity names in declaration materials into synonyms. This covers data of the same category with different expressions, and reduces missed recall cases. |
collectionIds | Collection IDs divided by oil product grade | This isolates knowledge base collections by oil product grade, avoids mixed recall of declaration data for different refined oil grades, and accurately matches the current declared oil product category. |
Three Common Mistakes
- Symptom: Retrieval results mix declaration materials for different oil product grades, returning files unrelated to the currently declared oil product. Cause: Failed to configure
collectionIdsto isolate knowledge base collections by oil product grade, leading to mixed recall of cross-category data. - Symptom: The large model reply does not display compliance inspection photos embedded in the knowledge base, only returning text content. Cause: Did not explicitly require displaying image links or embedding identifiers in the prompt, and the default setting only extracts text content.
- Symptom: The retrieval results output by the large model have custom modifications, and the original text format is not directly returned. Cause: Did not explicitly limit the prompt to only output the original knowledge base text, allowing the large model to make free revisions.
How to Confirm the Configuration Is Correct
- Upload a single specified-grade refined oil retail declaration form, perform retrieval, and check that the recall results only include relevant files and ledger data for that grade.
- Trigger a retrieval task for monthly filing files, verify that the update time of the returned results matches the current month.
- Configure the prompt to require only returning the original knowledge base text, perform retrieval, and check that the large model output fully matches the original knowledge base text content.
- After enabling
usingReRank, check whether the order of retrieval results matches the relevance to the current declaration topic.
The values provided here 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.