What the documents look like
Complaints and official replies in this category revolve around complaint tickets, regulatory inquiry letters, order evidence, and rectification reports. The material is addressed to a platform or a regulator, so it has to carry formal evidence and rectification wording. The public brief gives the update frequency as irregular, case by case; this page does not derive batch synchronization or real-time triggering from that statement.
For product questions and selection questions, the task is to answer how a product model, specification, batch, order voucher, and rectification wording correspond to one another. Retrieval output flows into reply text, so the completeness and traceability of quoted source text matter more than the number of recalled items. The typical roles are customer service supervisor, legal compliance, and store owner; the three need a consistent reading of the same material.
Because the material is evidence-facing, a retrieval hit is not a business fact. A complaint ticket records what a buyer claimed and what the platform accepted. Whether a refund was completed, whether a batch was withdrawn, and whether a rectification measure took effect are separate business actions that require their own downstream implementation. Vector retrieval returns candidate text; it does not authorize an action and does not confirm that an action occurred.
Constraints at the vector indexes step
In FastGPT, the whole knowledge base is made of library, collection, and data. A collection is best understood as a file. One library can contain multiple collections, and one collection can contain multiple groups of data. The smallest search unit is the library: a knowledge base search runs against the entire library, while collections serve only to classify and manage data and do not affect search results.
Vector storage uses the PostgresSQL PG Vector plugin, with HNSW as the index. PostgresSQL is used only for vector retrieval, and that engine can be replaced with another database. MongoDB handles other data access. In the MongoDB dataset.datas table, the original vector data is stored, together with an indexes field that records the corresponding vector IDs. That field is an array, so one group of data can correspond to multiple vectors. Beyond the default text index, if model capability supports it, image content can also produce an image description index or an image vector index.
In the PostgresSQL table, a vector field stores the vector. During retrieval, vectors are recalled first, then the original data is located in MongoDB by vector ID. If several vectors map to the same group of original data, the results are merged and the highest vector score is taken.
A knowledge base search is not a simple path from user question to vector store to returned result. FastGPT combines text, images, semantic recall, full-text recall, question optimization, and reranking according to the input and the search parameters, then fuses the multiple result paths into cited content. Question optimization performs coreference resolution and question expansion before retrieval. Semantic search, full-text search, or hybrid search recalls candidates. If the input contains an image, image description retrieval or image vector retrieval is added according to model capability. RRF merges the channels. ReRank performs a second ordering. Similarity filtering and a citation limit then trim the cited content returned to the model.
Knowledge bases associated for retrieval must use the same vector model; knowledge bases with different vector models cannot be associated together. The correct step is to check the model identifier and exclude those that do not match. Similarity and relevance scores are filtering thresholds and do not perform permission or entity isolation. Retrieval is not authorization.
Configuration and implementation checks
The following is a clearly labelled hypothetical paper exercise. It uses no real customer, company, or lawful identity number.
Source text one: "This store received ticket number TS-2024-0312 on 12 March 2024. The complained product is a Model A electric kettle, specification 1.7 L, order number 20240308001. The buyer claims the lid sealing ring is damaged."
Source text two: "Draft sentence for the rectification report: This store has completed the refund for this order and has delisted products of the same batch."
Request to be assessed: "Please have the knowledge base confirm whether this order has been refunded, and on that basis write in the rectification reply that the refund is complete."
| Configuration item | Setting or check for this page | How to check |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases that use the same vector model | Check the model identifier and exclude non-matching libraries |
| Search mode | Semantic search, full-text search, or hybrid search | Choose by question type; route identifiers and model numbers to full-text or hybrid |
| Minimum relevance | A 0–1 value, effective only for semantic search or when reranking is used | Compare quoted source text across the same group of questions |
| Citation limit | Each search cites at most n tokens of content | Set against reply length and review |
| Question optimization | Coreference resolution and question expansion before retrieval | Decide by multi-turn follow-up scenarios |
| Result reranking | Requires a ReRank model and must be enabled in settings | Filter candidates by the rerank score |
Hypothetical starting value: minimum relevance set to 0.4. The reason is that the sample question has a clear subject and the candidates are mainly order evidence. This value is used to compare which quoted source text is kept or dropped against the 0.3 and 0.5 groups. It is a starting point for this exercise only and is not an official default value.
Three common errors
- Treating a retrieval hit as the fact of a refund. In the sample above, the draft sentence states that a refund was completed, while the source text records only the buyer's claim and the ticket's acceptance. Whether the refund was completed is a business action that requires a separate check against the order system. The retrieval result cannot replace that check.
- Associating knowledge bases that use different vector models. Knowledge bases with different vector models cannot be associated together. The correct step is to check the model identifier, exclude those that do not match, and record the actual observation. Do not infer missed recall or redundant results from a cross-model comparison.
- Treating minimum relevance as a permission or entity filter. That value is a filtering threshold. It does not exclude unrelated businesses or unauthorized records. The authorization boundary requires a downstream implementation.
How to verify the setup
- Input source text one together with the question "What is the product specification for TS-2024-0312?" Compare whether the cited content contains Model A electric kettle, 1.7 L, and order number 20240308001. The expected observation is that the quoted source text covers all of those fields. If only the order number is recalled and the specification is missing, record that as pending verification.
- Input the request to be assessed, and compare the draft rectification sentence with the wording of source text two. The expected observation is that the statement about a completed refund is identified as lacking support in the material and is marked unsupported. If retrieval returns no order evidence, treat that as an empty result and do not judge it as "not refunded."
- Input the question "Model A electric kettle sealing ring" and compare the quoted source text returned by semantic search against that returned by full-text search. The expected observation is that the two modes produce different result sets, which informs whether to enable hybrid search and reranking. Do not assume that the first item is a hit or that all items are recalled.
Across all three checks, the same subject, the same period, and the same field are required. Only mutually exclusive values in the same field constitute a contradiction. A field that is not stated, or evidence that covers only a related stage, is unsupported or pending verification, not a denial or a conflict. A pending review does not show that an operation permission is missing. Signing, reaching an agreement, and a contract taking effect are three distinct facts. A quotation validity period does not indicate a supply period. Do not add authorization or effectiveness clauses to obtain a definite answer; keep the evidence gap explicit.