What the documents look like
For marketplace advertising and on-site traffic work, the documents behind product and selection questions are advertising plans, keyword reports, audience package descriptions, and conversion attribution tables. These materials update by day or by hour. Their fields commonly include plan identifiers, keywords, bids, audience package markers, and attribution definitions. Fast updating means a parsed result has to point back to the source version. The same plan identifier can carry different bid values and attribution meanings on different dates, so a chunk that loses its version reference mixes periods together.
Selection questions mix terminology, identifiers, and natural-language phrasing. A buyer may ask about a plan number, or may ask in ordinary words which plan suits a summer campaign. Both routes have to reach the same source passage. When chunking splits a condition away from the content it constrains, the retrieved fragment arrives without its applicable premise. The plan number, the effective period, the daily budget, and the attribution definition can end up in separate chunks, and the answer then reads a fragment as if it were the whole statement.
The search unit is the library, not the collection. Collections classify data for management and do not affect search results. Within a library, one data item can map to several vectors, so a single source passage can be recalled through more than one index. This is the mechanism that lets terminology, identifiers, and natural-language phrasing each reach the same original content.
Constraints at the document chunking step
Segment by the source structure and keep document identity and version with the passage. Distinguish original data from retrieval indexes. In the stored structure, the original data record holds an indexes field listing the vector identifiers that correspond to it, and one data item can correspond to multiple vectors. During retrieval, vectors are recalled first, then the original content is located by vector identifier in the original data store, and results that map to the same original data item are merged with the highest vector score kept.
Use terminology, identifiers, and natural-language phrasing as separate sample queries. After source material changes, check the actual training and indexing state. A successful upload does not mean retrieval has updated. Semantic recall and relevance scoring affect ordering and filtering only. They do not perform permission or entity isolation, and a similarity setting does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization.
Libraries built on different vector models cannot be associated together. Check the identifier first, exclude the non-matching library, and record what is actually observed. Do not run a cross-model retrieval to compare missed recalls.
A hypothetical paper exercise sample, used only for paper checking and not representing any actual campaign or approval: the source text reads, "Plan identifier QH-2210 runs from 1 March 2025 to 31 March 2025, daily budget 800 yuan, keyword 'summer sun protection' bid 1.2 yuan. The attribution definition for this plan is 7 days after click." The draft sentence to be judged reads, "QH-2210 was attributed 7 days after click in March 2025, with a daily budget of 800 yuan." The paper check compares the draft sentence against the source text on subject, period, and field values.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid search | Identifiers and natural-language questions coexist; compare the semantic route and the full-text route recall differences |
| Reference limit | 2000 tokens for this exercise | A starting point for this trial only; compare against average segment length and adjust |
| Minimum relevance | 0.4 for this exercise | A starting point for this trial only; compare against the rerank score distribution |
| Question optimization | Enabled | Multi-turn follow-up and short-question scenarios; compare retrieval hits before and after enabling |
| Segmentation basis | By source structure | Keep conditions with the content they constrain and point back to the source version |
| Post-update recheck | Check training and indexing state | Compare the upload record against what is actually retrievable |
The reference limit is expressed in tokens, the documented unit, and is not a top-k item count. Minimum relevance takes effect only under semantic search or when result reranking is used, and it is a filter threshold rather than a final ordering rule. Question optimization adds one model call before retrieval and typically increases overall latency. RRF is a fusion process that combines multiple recall channels; it is not a toggle. Image retrieval requires an image input and a model that supports it, and reranking requires a rerank model enabled under its settings. The documented flow diagram lists optional paths, and not every path runs on every request.
Three common errors
- Treating a successful upload as retrieval having updated. The input is the updated source material and the same query; the comparison is between the upload record and the content actually recalled; the observable result is that the query still returns a fragment from the earlier version. The upload record alone does not show the training and indexing state.
- Treating a semantic recall score as a permission or entity filter. The input is a query that crosses business boundaries; the comparison is between returned content and the intended scope; the observable result is fragments outside the intended scope. A relevance score is not a category, entity, or permission filter.
- Writing RRF fusion as an enablement switch. The input is a configuration item list; the comparison is between the parameters the documentation lists and that list; the observable result is an entry for a switch the documentation does not describe. RRF is the merge step for hybrid results.
How to verify the setup
- Input the hypothetical source text and the draft sentence given above. Compare the draft sentence fields against the source on subject, period, and field values. The condition is satisfied when the attribution definition and the budget both have a corresponding expression in the source text. This is a paper expectation and has not been run.
- Input the identifier QH-2210 and a natural-language question. Compare the fragments recalled by the two routes. The condition is satisfied when the identifier route hits the precise fragment and the natural-language route hits a semantically close fragment, and both point back to the same source version. This is a paper expectation and has not been run.
- Input an empty-result downstream scenario, using the paper input [] to check the downstream design. Compare how the downstream handles an empty reference. The condition is satisfied when the downstream does not treat an empty reference as authorized or executed. This is a paper expectation and has not been run.
Across all three checks, missing evidence remains missing. 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, the value is unknown rather than absent. Any business verification, approval, database write, or external synchronization requires its own stated downstream implementation.