What the documents look like
Complaints and official replies in the e-commerce after-sales group differ from repair or replacement work because the audience is a platform or a regulator. The material has to reconstruct a transaction and support statements about procurement origin and stock status. For the inventory procurement and supplier records task, the documents include purchase orders, supplier qualification files, inbound records, return vouchers and correspondence. They come from the order system, customer service chat, the 12315 platform and the platform complaint center. Each case is assembled separately and updates are irregular.
In the vector model and index step, these materials enter the knowledge base as text chunks. One original data record can be linked to several indexes. Retrieval recalls vectors first, then uses the vector ID to find the original content in the dataset.datas table in MongoDB. Records that map to the same original data are merged and the highest score is kept. Knowledge bases that are associated for retrieval must use the same vector model, so the actual model and input type are checked before anything else.
Constraints at the vector indexes step
Supplier names, purchase batches, inbound quantities and return reasons sit alongside natural-language descriptions. Semantic retrieval suits the descriptions; full-text retrieval suits identifiers and proper nouns. Hybrid retrieval combines recall paths through RRF, and a ReRank model reorders candidate text and returns a score between 0 and 1. The minimum relevance value applies only under semantic retrieval or when result reranking is used. It is a filter threshold and does not provide permission or entity isolation.
Image retrieval depends on whether an image description or image vector was generated at ingestion, and on the capability of the model at query time. After a model or index change, the same group of questions is compared against the quoted source text; a vector score alone is not treated as a promise of quality.
Hypothetical drill sample
Source text for comparison: "On 10 March 2025, Supplier A sent Store B a quotation, quotation number BJ-20250310, valid until 10 April 2025. On 12 March 2025, Store B replied confirming purchase intent, but the two parties did not sign a procurement contract. On 20 March 2025, Store B took into stock a batch of goods labelled with Supplier A's brand, inbound receipt RK-20250320, quantity 200 units."
Draft sentence to be assessed: "Supplier A and Store B have reached a procurement partnership, and the contract is in effect."
Assessment: the material records only the issuance of a quotation, a reply of purchase intent and an inbound record. It does not record contract signing or effectiveness clauses. Under the same subject, same period and same field, the quotation validity date does not represent a supply period, and purchase intent does not represent contract effectiveness. The draft sentence is unsupported and the evidence gap must be preserved.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Note and comparison method |
|---|---|---|
| Vector model | Matches the actual model of the associated knowledge base | Check the actual model identifier and input type; exclude mismatches and do not run cross-model retrieval comparisons |
| Index content | Each chunk focuses on a single supplier or a single batch | Build indexes from complete, focused information; weigh retrieval coverage against index length using the sample |
| Search mode | Hybrid retrieval | Semantic and full-text recall are fused through RRF; identifier questions can lean on full-text |
| Minimum relevance | Hypothetical starting value 0.3 for this trial | Chosen because the sample question has a clear subject; compare with 0.5 to observe recall range changes |
| Reference cap | Hypothetical starting value 2000 tokens for this trial | Controlled in tokens; compare with 1000 to observe quoted content length |
| Image index | Generate an image description or image vector at ingestion | Confirm whether the vector model supports image input; check image clarity and subject |
Three common errors
- Treating retrieval as authorisation. A query for "Supplier A contract is in effect" may retrieve the quotation source text, but the quotation only evidences that a quotation was issued. It does not evidence contract signing or effectiveness. Retrieval results do not carry permission or entity isolation functions.
- Associating across vector models. When knowledge bases built on different vector models are associated together, checking the identifiers reveals the mismatch. The non-conforming base should be excluded and recorded; cross-model retrieval should not be run to compare missed recalls.
- Treating quotation validity as a supply period. The material states the quotation is valid until 10 April 2025. A draft sentence writing "supply period until 10 April 2025" changes the field: validity and supply period are different facts, and that sentence is unsupported.
How to verify the setup
- Enter the hypothetical sample source text and the draft sentence "Supplier A and Store B have reached a procurement partnership, and the contract is in effect." Compare the three facts recorded in the material (quotation, intent, inbound) with the draft sentence's claim of contract effectiveness. Expected observation: the draft sentence is unsupported, and the evidence gap is recorded as missing contract signing and effectiveness clauses.
- Enter "BJ-20250310" for full-text retrieval and compare whether the recalled content contains the quotation number as written. Expected observation: the quoted content contains that number and the quotation validity date. If it is not recalled, check the full-text retrieval configuration and the index content.
- Enter a clearly paper-based result of
[]and check whether the downstream chain still runs. Expected observation: when the quoted content length is 0, the output chain still runs, consistent with the node document. Do not claim that a fictional rule retrieval necessarily returns[].
The public technical facts checked for this feature (used to identify relevant content in the documents; the body is written independently and is not required to reproduce a fixed five-row table): [["Vector model", "Knowledge bases associated for retrieval must use the same vector model; check the actual model and input type first."], ["Original data and indexes", "One original data record can be linked to several indexes; check indexes and source content separately."], ["Index content", "Build indexes from complete, focused information; weigh retrieval coverage and index length against the sample."], ["Image index", "Image retrieval also depends on whether an image description or image vector was generated at ingestion, not only on the query side."], ["Version and result review", "After a model or index change, compare the same group of questions against the quoted source text; a vector score alone is not treated as a promise of quality."]]