What the documents look like
Wholesale stall trading depends on on-site bargaining and consignment shipping, so the records are fragmented and transaction-level. The main documents are purchase sheets, handwritten sales slips, stall lease contracts, and payment receipts. Handwritten sales slips often carry abbreviations, corrections, and missing fields. Stall lease contracts concentrate terms and carry a fixed term. Payment receipts make the amount and date the points to verify. Updates arrive daily and per order, and the same batch of goods can appear across several documents. Product questions here ask about specification, model, and the stall or purchasing scenario the item fits. A retrieval unit must point back to a specific document and version, because the same model number across different slips cannot otherwise be told apart.
Constraints at the document chunking step
Parsing keeps the original structure and document identity first. At chunking, conditional clauses, amounts, terms, and the items they constrain cannot be split apart. A handwritten sales slip's note and the goods line it belongs to should stay inside the same data group. In FastGPT, one data group can map to multiple indexes, so the same passage can carry a terminology index and a number index; a hit on either index recalls that data group. The library is the search unit, and a collection only classifies data, so it does not affect search results. Search results are fused through RRF across multiple recall paths, then reordered by ReRank. The minimum relevance value takes effect only in semantic search or when result reranking is used. Retrieval and relevance do not carry permission or entity isolation, and libraries built on different vector models cannot be associated together.
Hypothetical paper sample: the source text reads "On 12 March 2024, stall A issued a handwritten sales slip to buyer B, goods: stainless steel shelf, three-tier specification, quantity 20, unit price 45 yuan, payment not received, agreed consignment to market C." The proposed sentence under review is "This sales slip confirms payment settled and shipment may be arranged." The source states only that payment was not received, and the proposed sentence takes the opposite value on the same field, so it is unsupported.
Configuration and implementation checks
| Configuration or check item | Hypothetical trial starting value | Reason and comparison object |
|---|---|---|
| Search mode | Hybrid search | Model numbers and natural-language questions coexist; compare recall coverage against pure semantic search |
| Minimum relevance | 0.35 | Hypothetical starting value; compare with 0.2 to observe whether irrelevant fragments decrease |
| Reference limit | 1200 tokens | Hypothetical starting value; compare with 800 to observe whether conditional clauses are truncated |
| Question optimization | Enabled | Multi-turn follow-up scenario; compare coreference resolution against disabled |
| Result reranking | Enabled | Second-pass ordering when candidates are many; compare ordering stability against disabled |
| Chunk review | Check slip by slip | Manual action to confirm conditions and their constrained content are not split apart |
These values are hypothetical trial starting points, not official defaults. Image retrieval requires image input and model capability, and reranking requires a model enabled in settings; neither path necessarily runs on every request.
Three common errors
- Treating a successful upload as an updated retrieval index. Upload, training, and index status are separate facts; check the actual training status before searching.
- Treating minimum relevance as a permission or entity filter. That value only filters low-relevance results and does not exclude unrelated business or unauthorized records.
- Describing RRF as a switch. RRF is the result-fusion process of hybrid search, not an independent control item that can be turned on.
Corresponding checks: for the first, input the changed slip and the same search term, compare the upload record with the actual recalled content, and observe whether the recall still returns the old version. For the second, input a cross-stall query, compare the returned records with the target scope, and observe whether records outside the scope appear. For the third, input a hybrid search request, compare the returned results with single-path recall, and observe whether multi-path hits are merged.
How to verify the setup
- Verification one: input the hypothetical sample source text and the proposed sentence, compare "payment not received" with "payment settled" for the same subject, same period, and same field. The two values are mutually exclusive, so the proposed sentence is judged unsupported.
- Verification two: input the term "stainless steel shelf," a number-type question, and a natural-language question, then compare the three recall paths and observe whether the same data group can be hit through different indexes. First-position hits and full recall are not guaranteed.
- Verification three: input the paper value [] as an empty result, check whether the downstream design handles it as an empty reference, and record the actual observation. A pending review status does not demonstrate missing operation permission; signing, reaching cooperation, and contract effectiveness are three separate facts, and missing evidence stays pending verification.