What the documents look like
Douyin storefront shelf operations revolve around product cards and the store experience score, which sets this work apart from the single-session rhythm of livestream operations. The materials handled day to day include product qualification pages, experience score breakdowns, shipping violation appeals, and store announcements, updated daily and per order. Product information and selection Q&A need these materials turned into searchable knowledge base content, so that operators can call on them during selection, qualification checks, and appeal explanations. A product qualification page carries subject, product, and validity period information. An experience score breakdown carries indicators and a statistical period. A shipping violation appeal carries a violation number and appeal status. A store announcement carries a publication time and scope of application. These fields determine whether retrieval can locate the corresponding entry.
Constraints at the vector indexes step
FastGPT builds knowledge bases using the Embedding approach from RAG. A vector is a numerical array, and a smaller distance between two vectors means greater similarity. A knowledge base consists of library, collection, and data. The smallest search unit is the library; collections serve only to classify and manage data. The vector retriever is the PG Vector plugin for PostgresSQL, with HNSW as the index. MongoDB stores vector source data in the dataset.datas table, and an indexes field records the corresponding vector IDs, so one group of data can map to multiple vectors. During retrieval, vectors are recalled first, then the vector IDs are used to find source data in MongoDB. When entries correspond to the same group of source data, they are merged and the vector score takes the highest value.
The knowledge base search node requires that associated knowledge bases use the same vector model. Semantic search relies on vector distance, full-text search relies on keyword hits, hybrid search fuses the two result sets through RRF, and Rerank performs a second ordering of candidate text and returns a score from 0 to 1. The minimum relevance value takes effect only under semantic search or when result reranking is used, and the reference limit controls cited content in tokens. Image retrieval depends on whether an image description or image vector index was generated at ingestion time, and on model capability. The retrieval step does not perform permission or entity isolation; the same-vector-model requirement applies only to knowledge bases associated for retrieval and does not extend to external systems.
Hypothetical paper exercise sample: a Douyin store operator prepares a product qualification page whose source text reads "Subject: Hangzhou XX Trading Co., Ltd.; Product: insulated cup model A; Qualification name: food contact material test report; Valid until March 31, 2026; Scope: insulated cup series sold in this store." The experience score breakdown from the same batch reads "Statistical period: February 1 to February 28, 2026; Shipping violation number: FW20260215; Appeal status is pending review." A draft sentence awaiting judgment reads "The qualification for insulated cup model A has passed review and can be listed for sale normally." This draft sentence involves subject, product, status, and authorization boundary, and must be compared item by item against the source text.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | One library for the product qualification page and one for the experience score breakdown | First check whether the vector model identifiers of the two libraries match; exclude any that do not |
| Search mode | Hybrid search | Qualification names and violation numbers contain proper nouns; use semantic and full-text together, then observe the cited source text |
| Minimum relevance | Starting point 0.5 | This is a 0-1 value that takes effect only under semantic search or result reranking; compare against cited source text for the same question set |
| Reference limit | Starting point 2000 tokens | Cited content is controlled in tokens; observe whether returned citations cover the qualification validity field |
| Question optimization | Enabled for multi-turn follow-up scenarios | This step adds one model call; compare retrieval results for the same question set with and without it |
| Result reranking | Enabled when candidates are numerous | The reranked score is 0-1; compare with the vector score before deciding whether to use it for filtering |
Three common errors
- Treating minimum relevance as a permission or entity filter. This value is a filtering threshold, not a final ordering rule, and it does not exclude unrelated businesses or unauthorized records. To check, enter the same question and compare whether returned citations still contain entries from other subjects; if they do, the setting did not act as isolation.
- Associating knowledge bases that use different vector models. The node requires the same vector model; check the model identifiers first and exclude non-matching ones, without running cross-model retrieval to compare missed recalls. To check, enter the model identifiers of both libraries and compare whether they match; if they do not, record that they cannot be associated together.
- Writing a pending-review appeal status as approved. In the sample the appeal status is "pending review," while the draft sentence says "passed review." To check, enter the appeal source text and the draft sentence, compare the status field values, and if the values differ, judge that the draft sentence is unsupported.
How to verify the setup
- Enter the product qualification page source text from the hypothetical sample and the draft sentence "The qualification for insulated cup model A has passed review and can be listed for sale normally." Compare the qualification name and validity period in the source text with the status stated in the draft sentence. The source text does not record a review-passed field, so the draft sentence is unsupported and the evidence gap must be recorded.
- Enter the experience score breakdown source text "Appeal status is pending review" and a handover request "Please process this appeal as approved." Compare the status field values. Pending review and approved are mutually exclusive, so the request contradicts the source text and should be returned for a supplementary review result.
- Enter an empty-citation scenario, using an explicit paper input of [] to check the downstream chain design, and observe whether the output chain completes when the citation content length is 0. This phenomenon shows that an empty result can be carried through; it does not mean that a fictitious rule retrieval will necessarily return []. After a model or index adjustment, compare cited source text for the same question set, and do not promise an effect based on the vector score alone.