What the documents look like
Platform product title and detail compliance work deals with advertising truthfulness and qualification display rather than parcel delivery. The working materials are product title tables, detail page assets, advertising prohibited-word lists, and quality inspection reports. Sources include platform rule centers, advertising law compliance checklists, brand asset libraries, and inspection agency reports. The most time-consuming steps are checking absolute-word claims, comparing physical parameters, and replacing non-compliant assets. Errors can lead to product delisting, complaints, and administrative penalty risk. Materials update weekly and with each new listing.
In product material and selection Q&A, these four document types must be retrieved, cited, and handed to a human reviewer. Vector model and index configuration therefore determines whether cited content can cover titles, assets, prohibited words, and reports together. A title table carries short, structured entries where exact wording matters. A detail page asset carries longer promotional copy where semantic similarity helps. A prohibited-word list carries exact terms that keyword matching handles better. An inspection report carries identifiers and measured values that must be matched exactly. Retrieval and similarity scores do not perform permission or entity isolation, and they do not decide compliance.
Constraints at the vector indexes step
A knowledge base consists of libraries, collections, and data. The smallest search unit is the library; collections only classify data and do not affect search results. FastGPT uses the PostgresSQL PG Vector plugin as the vector retriever with an HNSW index, and MongoDB for other data storage. The dataset.datas table stores vector source data and holds an indexes field recording corresponding vector IDs as an array, so one data item can map to multiple vectors. During retrieval, vectors are recalled first, then the vector ID is used to find source content in MongoDB; when multiple vectors map to the same source item, results merge and the vector score takes the highest value.
Knowledge bases associated for retrieval must use the same vector model, so the actual model and input type are checked first. Index content should be complete and focused; retrieval coverage and index length are weighed against the sample. Image retrieval also depends on whether image description or image vector indexes were generated at ingestion time, not only on the query side. Hybrid retrieval merges semantic and full-text results through RRF. Result reranking uses a ReRank model for secondary ordering. The minimum relevance threshold takes effect only in semantic retrieval or when result reranking is used. A relevance score is not a category, entity, or permission filter.
Configuration and implementation checks
Configuration below centers on the hypothetical sample and lists only documented settings or implementation checks. Numeric values are starting points for this hypothetical exercise, used for comparison against actual values in the sample, and are not official defaults.
| Configuration item | Hypothetical starting point | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Four libraries: product title table, detail page assets, advertising prohibited-word list, quality inspection report | All four material types fall within this task's retrieval scope; confirm they share the same vector model |
| Search mode | Hybrid retrieval | Prohibited words and report numbers are exact queries; titles and assets are semantic queries; RRF merges both paths |
| Minimum relevance | Hypothetical starting point 0.35 | Compared against semantic scores of prohibited-word hit fragments in the sample; results below this value do not enter citations |
| Citation limit | Hypothetical starting point 2000 tokens | Combined fragment length across the four material types in the sample falls near this range; compared against total returned citation length |
| Result reranking | ReRank enabled | Candidate fragments come from four libraries; rerank scores provide secondary ordering and filtering |
| Question optimization | Enabled | The draft sentence contains coreference and compliance-judgment intent; coreference resolution and question expansion run before retrieval |
Three common errors
- Treating vector similarity as a compliance determination. Observable error: after retrieving the "absolutely silent" fragment, the output states "meets requirements." Input is the prohibited-word list text and the draft sentence above. Comparison object is whether "absolute" appears in the prohibited-word list as a banned expression. Observable result is whether the retrieved fragment and the draft sentence both contain that term. A similarity score does not perform compliance judgment; a human compares against the prohibited-word list.
- Associating libraries that use different vector models. Observable error: libraries built on different vector models are selected into one retrieval. Input is the model identifier of each of the four libraries. Comparison object is whether the declared vector models match. Observable result is that mismatched libraries are excluded and recorded, and no cross-model retrieval comparison is executed.
- Treating a pending review request as approved. Observable error: the draft sentence sits in pending compliance review, yet the output states "can be listed directly." Input is the request status field and the draft sentence. Comparison object is whether any signature or approval record exists. Observable result is that a pending-only status is judged unsupported, and the evidence gap is preserved.
How to verify the setup
- Input the three material sentences and the draft sentence from the hypothetical sample above. Comparison object is the prohibited expressions "lowest price" and "absolute" in the prohibited-word list against the corresponding terms in the draft sentence. Observable result is that retrieved citations cover both the prohibited-word list fragment and the detail page asset fragment, and no compliance conclusion is output. Expectation is that citations are presented in full and compliance judgment is left to a human. If retrieval returns an empty result, a paper input of [] can be used to check downstream design; do not claim that a fictional rule search necessarily returns an empty array.
- Input the model identifiers of the four libraries. Comparison object is whether the declared vector models match. Observable result is that matching libraries enter retrieval and mismatched libraries are excluded and recorded. Expectation is that no cross-model retrieval comparison is executed, and no inference about missed recall or redundancy is drawn from the exclusion.
- Input the request status field. Comparison object is whether signature, approval, or contract-effective records exist. Observable result is that a pending-only status is judged unsupported. Expectation is that no authorization or effective clause is fabricated and the evidence gap is explicitly preserved.
Hypothetical exercise sample. A platform store operator submits a product material review request. The material text has three sentences. Sentence one: "The product title is 'household silent wall-breaking machine, large capacity'." Sentence two: "The detail page asset states 'lowest price on the entire internet, absolutely silent'." Sentence three: "The advertising prohibited-word list specifies that 'lowest price' and 'absolute' are prohibited expressions; quality inspection report number QC-2024-0871 shows measured noise of 78 decibels." The draft sentence to be assessed is: "The product title and detail page asset both meet platform compliance requirements and can be listed directly." The request status is pending compliance review, with no signature or approval obtained.