What this category’s data looks like
Coal transportation and sales ledger data has four primary sources: coal mine weigh tickets, road or railway waybills, purchase and sales contracts, and monthly settlement reports. Data update frequencies are categorized as daily (weigh tickets, real-time waybills), weekly (settlement detail ledgers), and monthly (full summary reports). Most individual documents are structured tables. They include fields such as mine site code, rail car or license plate number, shipment tonnage, transportation mileage, receiving unit, settlement unit price, and freight amount. Supported units include tons, yuan, trips, and kilometers. Some documents are available as scanned image files.
Constraints imposed on vector models and indexing
Mixed-format document sources require the indexing pipeline to support combined capabilities: scanned image OCR, structured Excel extraction, and PDF text parsing. Daily batch-generated weigh tickets and waybills create high-frequency indexing write requests. Single-batch document count must be limited to avoid overloading index nodes. Many fields contain numerical values with associated units. Vector models must accurately encode such structured text with units. This prevents misclassification of semantically similar fields with different units as related. Long-text structured monthly summary reports require indexing chunking strategies that preserve the integrity of table-style long text.
Configuration settings
This configuration is optimized for FastGPT v4.17.0, tailored for coal transportation and sales ledger scenarios.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Aligns with the row length of structured tables in coal transportation and sales documents, avoids truncating critical numerical and associated fields |
| Overlap Length | 100–150 characters | Covers header and data associations across chunks, prevents loss of contextual links after chunking |
| Similarity Threshold | 0.72–0.78 | Matches the semantic distribution of coal transportation and sales documents, differentiates waybills from the same mine site across different batches, avoids incorrect matches |
| Single Search Limit | Top 8 results | Balances recall accuracy and response speed for monthly summary reports, covers the typical number of associated documents |
defaultToken | 4096 token | Aligns with the text length of individual settlement reports, ensures full loading of summary data |
maxToken | 8192 token | Covers full contextual requirements for ledger summary and report generation, avoids truncating critical total information |
Three common errors
- Attempting to use multiple vector models in the same knowledge base, leading to semantic drift in recall results. Reason: Different vector models have varying encoding logic for numerical fields with units in coal transportation and sales documents. Mixing models breaks semantic consistency.
- Directly integrating a graphRAG tool for index construction, leading to incorrect breakdown of structured field associations. Reason: FastGPT’s default indexing is designed for flat entities, and does not align with graphRAG’s relational indexing logic. Direct integration loses structured associations in ledgers.
- Enabling automatic supplementary indexing without proper configuration leads to extensive garbled text in OCR results for scanned documents, causing index recall failures. Reason: OCR parameters tailored for coal transportation and sales documents are not configured. Critical information such as weigh scale numbers and rail car IDs in scanned documents cannot be accurately identified.
How to verify correct configuration
- Upload a single weigh ticket document, review the chunked content after index generation, confirm chunk length and overlap length match preset settings, and no critical fields are truncated.
- Input a query statement containing a mine site name and tonnage, check the similarity scores of recall results, confirm no incorrect or missed matches after adjusting the threshold.
- Batch upload 3 to 5 waybills and monthly reports from different dates, review index generation status logs, confirm single batch processing does not trigger node load alerts.
- Call the ledger summary and report generation interface, review the returned context length, confirm the
maxTokenconfiguration covers the text content of all associated documents.
The values provided here are common starting points, and should be evaluated 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.