What the documents look like
Fresh-food supplier claims pursue responsibility for quality defects at goods receipt, and they stay separate from handling natural shrinkage inside stores. The working materials come from the quality inspection system, WMS receiving records, supplier contracts and finance reconciliation. Day-to-day documents include quality exception notes, photo evidence, deduction notices and claim reconciliation statements.
Two update rhythms shape this material. Each arriving batch produces receiving records, exception photos and a deduction basis that need to be grouped by batch. Monthly reconciliation produces the longer statements and contract clauses. Lengths vary widely across the set: an exception note description is short, while a reconciliation statement and contract clause run long. Photo evidence adds a further question, because whether a photo can be retrieved depends on whether a corresponding index was created at the time the material was stored.
For the vector model and index step, these traits create specific constraints. Vector search first relies on a vector model to convert text or images into numeric arrays, where a smaller distance means greater similarity. Knowledge bases that are associated for retrieval must use the same vector model, so the actual model identifier and input type are checked first. Libraries built on different models cannot be associated together; the identifier is verified and non-matching libraries are excluded. A single source data block can be linked to multiple indexes, and index content is checked separately from the original content. Index content is built from complete and focused information, and retrieval coverage against index length is weighed against sample material. Whether an image can be retrieved depends on whether an image description index or an image vector index was generated at storage time, and not only on whether an image was uploaded at query time. Semantic search suits natural-language questions, full-text search suits numbers, models and proper nouns, hybrid search fuses multiple retrieval paths through RRF, and result reranking re-sorts candidate text. Similarity scores and the minimum relevance value act only on ordering and filtering; they do not perform permission or entity isolation, and retrieval itself does not constitute authorization.
Hypothetical paper sample: a supplier contract states that acceptance of arrival batch A is based on quality inspection system records. Quality exception note QA-001 records that leafy vegetables arriving on 10 March 2025 showed compression damage. Deduction notice CN-001 proposes a deduction on that basis. A draft sentence reads: "QA-001 has confirmed full cargo damage responsibility for this batch, and the deduction has been executed." Checking compares the draft sentence against the exception note text: the note records the compression damage fact but does not record a responsibility conclusion, and the deduction notice remains at the proposal stage with no recorded execution. The draft sentence exceeds what the source text supports and is unsupported.
Constraints at the vector indexes step
The retrieval unit is the library, and a library can contain multiple collections while a collection contains multiple data groups. A search runs against the whole library; collections serve data classification and do not affect search results. In the MongoDB dataset.datas table, the vector source data is stored, and an indexes field records the corresponding vector IDs as an array, so one data group can map to multiple vectors. In PostgresSQL, a vector field stores the vectors, and retrieval recalls vectors first, then looks up the original content in MongoDB by vector ID. Where entries map to the same source data group, they are merged and the highest vector score is kept.
Multi-vector mapping exists because content length and semantic richness conflict within one vector. Mapping one data group to several vectors preserves both completeness and semantic richness. When a longer text is given several vectors, recall of any one vector brings the data group back. This means data block precision can be raised by annotating blocks.
The overall retrieval path is not a single user question to vector store to result. Depending on input content and search parameters, text, images, semantic recall, full-text recall, question optimization and reranking are combined, and multi-path results are fused into reference content. Question optimization performs coreference resolution and question expansion before retrieval, adding a model call and typically increasing latency. Image retrieval is not a separate capability outside the knowledge base; it adds an image input path to the existing search chain. A knowledge base search node can be added repeatedly, has external input, static configuration and triggered execution, and outputs reference content as an array whose length can be zero, so the output path completes even when nothing is found.
Configuration and implementation checks
The configuration below centers on sample checking and lists only documented settings or implementation checks relevant to this feature and task. The numeric values are hypothetical starting points for this paper comparison and do not represent official defaults.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Vector model | Record the actual model identifier | Compare character by character against the model identifier of each associated knowledge base; exclude any library that does not match |
| Search mode | Hybrid search | Compare against semantic search and full-text search on the same question set, checking the returned source text |
| Reference limit | Hypothetically set to 800 tokens | Compare reference count and source completeness against a hypothetical 500-token setting |
| Minimum relevance | Hypothetically set to 0.4 | Compare returned count and proportion of irrelevant fragments against hypothetical 0.2 and 0.6 settings |
| Index content | Build separately for exception notes and deduction notices | Check against source text sentence by sentence, looking for missing subject, batch or status |
| Image index | Check whether one was generated at storage time | Check whether the source data behind the photo evidence records an image description or image vector index |
Three common errors
- Associating knowledge bases built on different models. The input is two libraries with different model identifiers; compare the model identifier fields and observe whether both are still selected into the same retrieval. When model identifiers differ, the libraries should not be associated together.
- Inferring authorization from a similarity score. The input is a low-relevance deduction notice belonging to a different supplier; compare the returned content against the current supplier scope and observe whether it enters the references because of a high score. Similarity filtering does not perform entity isolation.
- Treating a deduction notice as executed. The input is the draft sentence "the deduction has been executed"; compare it against the status field of the deduction notice text and observe whether the source records an execution result. Where the source only records that a deduction was proposed, the sentence is unsupported.
How to verify the setup
- Check one: input exception note number QA-001 and the draft sentence from the hypothetical sample, compare the draft sentence against the subject, batch and status fields of the exception note text, and observe whether the draft sentence presents a responsibility conclusion the source does not record. Its appearance means the condition is not met.
- Check two: input the same question set once with semantic search and once with full-text search, compare the hits on numbers QA-001 and CN-001 in the returned source text, and observe whether full-text search hits the numbers more consistently. Where number hits are unstable, record the actual observation and keep the point open for confirmation.
- Check three: input an empty-result path, compare whether the reference content output is an array of length zero, and observe whether the downstream design accepts an empty array. This check verifies empty-result handling only and does not infer that any invented rule necessarily returns an empty array.
Operational handover must keep material, action and request authorization states distinct. Retrieval and forms are not authorization execution: a pending request is not approval, an offered option is not execution, and an unsigned termination is not a completed termination. Where a document field is not provided, it remains unknown.