What the documents look like
Fresh-food supplier claims pursue responsibility for quality at arrival. They do not address ordinary in-store shrinkage. The working materials come from quality inspection systems, WMS receiving records, supplier contracts and financial reconciliation. The documents handled day to day are quality exception forms, photographic evidence, deduction notices and claim reconciliation statements. Each arrival batch produces quality exception records; each month produces claim reconciliation.
These materials carry batch identifiers, supplier identities, receiving status and deduction amounts. A quality exception form and its photographs belong to the same batch. A deduction notice and a reconciliation statement share an amount, a period and a supplier subject. That structure is what document parsing and chunking must preserve, because a retrieved fragment that loses the batch, the supplier or the approval state cannot support a product question about that batch.
The brief for this category is marked as broadly accurate rather than verified. Data update frequency is per batch and monthly. Where a document field is not provided, it remains unknown.
Constraints at the document chunking step
Chunking decides what a retrieval hit actually contains. If a condition, an explanatory note and the clause it constrains are split apart, the retrieved fragment lacks the subject or the validity scope needed to judge it. In FastGPT, one set of source data can map to multiple indexes, and a hit on any one of those indexes recalls that set of data. Full-text retrieval suits identifiers, model numbers and proper nouns; semantic retrieval suits natural-language questions. A relevance score is a similarity measure, not a category, entity or permission filter, and retrieval is not authorization.
Hypothetical paper exercise sample. The following is invented for this page and was not taken from real records.
Source material for checking, two sentences: "Supplier batch of spinach arrived 10 March 2025. Quality inspection record: leaf yellow-spot proportion approximately 8%, above the 5% ceiling agreed in Annex 2 of the contract." A second source sentence: "WMS receiving note marked partial rejection. Deduction notice draft: proposed deduction of 1200 yuan, pending quality supervisor approval."
Draft sentence to be judged: "The deduction of 1200 yuan for this batch of spinach has been completed."
The source material shows a draft deduction notice awaiting approval. It does not record completed approval or executed deduction. The draft sentence is unsupported by the source wording.
Configuration and implementation checks
| Check item | Relation to this page's documents | Hypothetical starting value and comparison method |
|---|---|---|
| Search mode | Quality exception forms carry batch and proportion; deduction notices carry amount and state | Use "10 March spinach yellow-spot proportion" as a semantic retrieval input and "deduction notice draft" as a full-text retrieval input; compare whether the recalled fragments retain subject and state |
| Reference limit | Reconciliation statements and exception forms differ greatly in length | Start at a token amount covering one exception form plus one reconciliation summary paragraph; compare with a value covering only a single short sentence, and observe whether conditional clauses are truncated |
| Minimum relevance | Takes effect under semantic retrieval or result reranking | Start at 0.5 and compare with 0.3; observe whether low-relevance fragments still enter the reference content |
| Question optimization | Coreference is unclear in a follow-up such as "the deduction state of this batch" | With it enabled, compare whether the completed result contains batch and supplier subject, against the result with it disabled |
| Image index | Whether photographic evidence generated a corresponding index at ingestion | Check the ingestion record for an image description or image vector index for this batch's photographs; compare with retrieval over text-only material |
| Re-check after update | A deduction notice moving from draft to approved requires re-checking | After the change, check actual training and index state, and compare with retrieval results before the change |
RRF is the fusion step for mixed retrieval results, not a standalone switch. Image retrieval requires corresponding image input and model capability. Reranking requires a model and must be enabled by setting.
Three common errors
- Treating "deduction notice draft, pending quality supervisor approval" as a completed deduction. Input: the draft sentence above together with the source material. Comparison: the state field in the source material. Observable result: the draft sentence's claim of completed deduction has no corresponding record in the source, so it is unsupported.
- Treating a retrieval hit as authorized execution. Input: one retrieval request for this batch. Comparison: the approval state of the deduction notice. Observable result: the returned fragment does not change the draft state; authorization still requires a downstream approval implementation.
- Inferring missed recall after associating libraries that use different vector models. Input: two library identifiers each using a different vector model. Comparison: the model identifier of each library. Observable result: when the identifiers differ, shared association must be excluded, and the actual retrieval behaviour remains unknown.
How to verify the setup
- Use the hypothetical source material and the draft sentence as input. Compare the draft sentence's claim of completed deduction against the source's draft-pending-approval state. The condition is met when the draft sentence is judged unsupported.
- Retrieve with "10 March spinach yellow-spot proportion" and with "deduction notice draft" separately. Compare whether the recalled fragments retain batch, supplier and state fields. If a fragment lacks subject or state, record it as pending improvement.
- For the empty-result case, use the explicit paper input
[]to check the downstream design. Compare whether the downstream can handle empty references. Do not assert that a fictional rule makes retrieval return[].
All three verification steps are paper checks. No retrieval or approval was actually executed. A pending review does not demonstrate absence of operating authority. Signing, reaching cooperation and contract effectiveness are three distinct facts. A quotation validity period does not indicate a supply period.