What the documents look like
Presale group buying collects orders first and organizes procurement and fulfillment afterward. Inventory purchasing and supplier materials move through the presale product table, the consolidated order result, the procurement plan, and the cut-off notice. The subject, object, status, and validity range recorded in those materials determine whether vector retrieval can recall the correct fragment.
These materials update daily and per session. Data in a retrieval library must already be trained before it can be recalled. Supplier records and procurement plans frequently carry exact fields such as identifiers, specifications, and quotation conditions. Semantic retrieval depends on how the index is written for that kind of content, while full-text retrieval addresses identifiers and proper nouns more directly.
Constraints at the vector indexes step
A knowledge base search operates at the library as its smallest unit; a collection serves only as a classification layer. Vector retrieval runs on PostgresSQL with the PG Vector extension, using an HNSW index. MongoDB stores the original vector data in the dataset.datas table, and the indexes field records the corresponding vector IDs, so one data record can map to multiple vectors. Libraries linked for retrieval must share the same vector model. Check the actual model and input type first; libraries built on different vector models cannot be linked together, so identify the model and exclude mismatches.
Index content should be complete and focused. Retrieval coverage and index length need to be weighed against the sample at hand. Whether an image can be retrieved depends on whether an image description index or an image vector index was generated at ingestion time; uploading an image on the query side does not create an ingestion index. RRF is the fusion step for mixed retrieval results, not a toggle. ReRank requires a model and must be enabled through its settings. Semantic recall and relevance scores carry no permission or entity isolation; retrieval does not constitute authorization.
Configuration and implementation checks
Hypothetical paper exercise sample: the source material reads "Supplier A, quotation number CG-2026-0417, quotation valid until 31 May 2026, supplied category leafy vegetables, current status pending review." The sentence to be checked reads "Supplier A already holds a valid leafy vegetable supply qualification." That sentence shares subject and category with the source, but the source status field says pending review and states no effective qualification. The sentence is unsupported and the evidence gap must be recorded.
| Configuration item | Relevant setting or check on this page | Basis and limitation |
|---|---|---|
| Linked knowledge base | Select one or more knowledge bases with the same vector model | Libraries on different models cannot be linked; verify the model identifier first |
| Vector model and index | Verify the actual vector model, HNSW index, and input type | Retrieval conditions do not hold when model and input type do not match |
| Index content | Build indexes for procurement plans and cut-off notices from complete, focused information | Index length and retrieval coverage are weighed against the sample |
| Image index | Check whether an image description or image vector index was generated at ingestion | Uploading an image only on the query side does not create an ingestion index |
| Search mode | Identifier fields can use full-text or hybrid retrieval | RRF is a fusion step, not a switch |
| Reference limit and minimum relevance | Reference limit is counted in tokens; minimum relevance applies only to semantic retrieval or reranking | Relevance is a filter threshold and carries no permission isolation |
Three common errors
- Treating pending review as already effective. In the sample, "current status pending review" and the draft sentence "already holds a valid leafy vegetable supply qualification" are the same field with different facts. Pending review does not show the qualification is effective, and it does not show that an operating permission is missing.
- Treating a query-side image upload as an ingestion index. Image retrieval also depends on whether an image description index or image vector index was generated at ingestion. Uploading an image alone cannot make that image recallable.
- Forcing libraries on different vector models into a shared link. Linked retrieval requires the same vector model. When model identifiers differ, exclude the library and record it; do not run a cross-model retrieval to compare missed recalls.
How to verify the setup
- Enter the sample source text and the draft sentence, then compare the status field value for Supplier A. If the source says pending review while the draft sentence claims an effective qualification, judge the sentence unsupported and record the gap. The sentence is supported only when the source carries an effective qualification value in the same field.
- Enter a query containing quotation number CG-2026-0417 and compare whether the returned references contain that identifier from the source. Identifier queries use full-text or hybrid retrieval; observe whether it is hit, and record a miss as pending verification without guaranteeing a first-position hit.
- Enter an image query that points to an image with no generated image index, then compare the ingestion index record with the returned references. If ingestion produced no image description or image vector index, record that path as unavailable and do not claim that an unrelated query must be empty. An empty result can carry an explicit paper input of [] to check downstream design; do not claim that a fabricated rule retrieval must return [].
After a model or index change, compare references against the same set of questions and check the source text, rather than promising an outcome from a vector score alone.