What the data for this category looks like
Coal transportation and sales business data mainly comes from five categories: waybills, purchase and sales contracts, settlement vouchers, monthly transportation plans, and safety and compliance records. Data sources include offline scanned paper documents, files exported from online transportation systems, and structured field data synced from ERP systems. Update cycles vary significantly: waybills are generated in real time with each batch of transportation, monthly plans are updated monthly, and contracts are updated when signed. Document structure primarily uses structured fields, including goods name, tonnage, departure station, arrival station, rail car number, settlement unit price, and other fields. Units include tons, yuan per ton, trips, and similar units, alongside unstructured contract clauses and transportation instruction text.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The multi-type data and differentiated update cycles of coal transportation and sales create multiple constraints on the retrieval and recall process. Mixed-format documents (scanned PDF files, Excel structured tables, plain-text contracts) require the retrieval system to support both OCR recognition and structured field matching. Without this support, the retrieval system will miss valid content from scanned documents. Waybill data updated in real time at high frequency increases the frequency of incremental index synchronization. Improper configuration can lead to delayed recalled data. Features with multiple fields and units can also cause unit mismatch errors in single-semantic retrieval, requiring targeted field matching logic.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.78 | Adapts to the structured and mixed semantic features of coal transportation and sales documents, balances recall accuracy and coverage, avoids introducing irrelevant results when set too low, and prevents missing valid business data when set too high |
limit | Top 12 entries | Covers the reference data volume required for drafting a single coal transportation and sales document, avoids overly long contexts exceeding model processing limits, and aligns with the default context limit of FastGPT v4.17.0 |
searchMode | hybrid | Matches both full-text semantics and structured fields, adapting to the mixed content of structured fields such as goods name and rail car number alongside contract clauses in coal transportation and sales documents |
usingReRank | Enabled | Reorders initial recall results by relevance, corrects semantic matching deviations of structured fields from single vector retrieval, and improves result practicality |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands industry term queries related to coal transportation and sales, such as associating "ton" with common industry expressions, improving recall coverage across terms |
collectionIds | Exclusive coal transportation and sales business collection ID | Isolates business data, avoids interference from cross-category retrieval, and ensures recalled results only come from the target business knowledge base |
Three common mistakes
- Scenario: Empty results are returned when scanning a PDF of a coal transportation waybill. Cause: OCR recognition configuration is not enabled, only the text layer of the PDF is recognized, and document content in images is not extracted.
- Scenario: Structured fields are truncated or semantically broken after knowledge base chunking. Cause: Confusing character and token units for chunk length, failing to adjust chunking rules based on the average length of coal transportation and sales documents.
- Scenario: Retrieval returns a large number of duplicated waybill settlement data. Cause: Index deduplication configuration is not enabled, and incremental index optimization is not performed for high-frequency repeated business documents, leading to multiple recalls of the same data.
How to confirm the configuration is correctly set
- Upload a scanned PDF of a coal transportation waybill, initiate a query containing "rail car number", and check whether the retrieval results include the document content extracted via OCR.
- Initiate a query containing "settlement unit price", review the order of returned results, confirm that documents with high relevance are ranked first, and verify that the
usingReRankconfiguration is active. - Check the number of returned retrieval results, confirm that it matches the value set in the
limitconfiguration, and verify that the retrieval quantity limit is active. - Check the field matching status of retrieval results, confirm that structured fields such as "departure station" and "arrival station" are correctly identified, and verify that the hybrid configuration of
searchModeis active.
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.