What the documents look like
Procurement documents in this category come from wholesale market quotations, supplier quotes, store sales and testing reports. The working set is made up of purchase orders, acceptance records, pesticide-residue test reports and loss records. Purchase orders and acceptance records are produced with each delivery, while test reports are issued per batch, so the update rhythm is daily or per batch.
The structures differ in ways that matter at the chunking step. A purchase order carries supplier name, item name, quantity, unit price and date. An acceptance record carries grade determination, received quantity and the accepting person. A test report carries sample name, test item, result and report number. A loss record carries the written-off item name, quantity and reason. These fields are what a chunk must keep available when it has to point back to the original document and its version.
Constraints at the document chunking step
Chunking has to preserve the document identity and version that allow a retrieved fragment to be traced back to its source. On an acceptance record, the grade determination and the received quantity belong to one business fact; a condition and the content it constrains should not be split apart. One set of source data can map to several indexes, and the source data and the retrieval indexes live in separate storage. At search time, vectors are recalled first, then the source data is fetched by vector ID; when several vectors resolve to the same source data group, the results are merged and the highest vector score is kept.
The retrieval library has to match the vector model, and this requirement stays inside this feature. It does not extend to unrelated external systems. Similarity and relevance serve ordering and filtering only; they do not perform permission checks or entity isolation.
The following is a hypothetical paper exercise, used to show how checking works. It involves no real customer and no real identifier.
Source text one: Supplier A issued a quotation on 10 March 2025 for Red Fuji apples, grade one, unit price six yuan per kilogram, valid until 17 March 2025.
Source text two: The store issued an acceptance record on 12 March 2025 against purchase order number CG-20250310-01, recording eighty kilograms of grade-one Red Fuji apples received, signed by the accepting person Li Si.
Draft sentence to be judged: Supplier A had completed delivery of Red Fuji apples to the store by 12 March 2025.
The evidence needed for this judgment: the quotation states a unit price and a validity period, and states no delivery period; the acceptance record states a received quantity and an accepting signature, and states no contract effectiveness or completion of delivery. The claim of completed delivery is not supported by the available material and remains pending verification.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Search mode | Hybrid search | Purchase order numbers and item names need exact hits, while natural-language questions need semantic recall; both are used together |
| Minimum relevance | Hypothetical starting point 0.5 | This value takes effect only in semantic search or with result reranking; adjust after comparing against the score distribution actually returned |
| Reference limit | Hypothetical starting point 2000 tokens | The limit controls referenced content by tokens; compare against the actual length of content returned in a single search |
| Question optimization | Hypothetical on | Follow-up questions often lack a subject; compare against results with it off |
| Result reranking | Hypothetical on | When candidates are numerous, rerank scores are used for filtering; compare against ordering by vector score alone |
| Image retrieval | Hypothetical off | Current material is mainly text; first confirm whether the vector model supports image input |
Three common errors
- Treating the quotation validity period as a delivery period. The quotation states a unit price and a validity period, and states no delivery period. The check compares the same supplier, the same period and the same field: quotation validity and delivery period are different fields and cannot substitute for each other.
- Treating an acceptance signature as contract effectiveness or completed delivery. The acceptance record states a received quantity and an accepting signature; signing, reaching cooperation and contract effectiveness are three distinct facts. The check confirms whether the material states an effectiveness clause, and keeps the evidence gap when it does not.
- Treating a similarity score as a permission or entity-isolation mechanism. Similarity and relevance serve ordering and filtering only; retrieval does not perform authorization judgments. The check confirms whether the access scope is already bounded by the target retrieval scope, and does not infer record accessibility from a score.
How to verify the setup
- Enter the quotation text and the draft sentence from the hypothetical exercise, and compare the claimed completed delivery against the unit price and validity fields stated in the source. If the source states no delivery period and no effectiveness clause, the draft sentence is unsupported and should be recorded as pending verification.
- Enter the acceptance record text and purchase order number CG-20250310-01 from the hypothetical exercise, and compare the received quantity, the accepting person and the correspondence to the purchase order. If the number matches and the quantity field corresponds, the acceptance fact holds; if the number does not match or the quantity field is missing, the correspondence is unsupported.
- Enter the supplier name and item name from the hypothetical exercise, searching once by number and once in natural language. If the number search hits the purchase order and acceptance record, and the natural-language search hits the quotation and loss record, the effect of chunking on the actual question can be observed; if either path returns nothing, record that path as empty and check the index state, without inferring other causes.
After material changes, verify the actual training and index state; a successful upload is not evidence that retrieval has updated. When a search returns an empty result, the downstream design can be checked against a clearly stated paper input of an empty list, without claiming that any invented rule returns an empty list. Libraries built on different vector models cannot be associated together; check the identifiers, exclude those that do not match, and record what is actually observed rather than performing a cross-model search to compare missed recall. The optional paths in the overall flow diagram do not all run on every request: image retrieval requires image input and a capable model, and reranking requires a model and its configured enablement.