What the documents look like
Commission settlement happens after a sale closes. The records to reconcile are valid promotion orders and the commission deductions that follow after-sales activity. Day-to-day materials include commission setting records, settlement bills, refund deduction details, and invoices. These materials reference each other across fields: a bill line cites the commission rate from a setting record, and a deduction detail cites a refund order number and a settlement period. Product material and selection Q&A must answer which materials contain a given commission rule, a given deduction basis, or a given invoicing treatment, and whether those materials fall inside the same searchable knowledge base scope. Materials update weekly or monthly. The retrieval side only needs to confirm the currently ingested version; the update cadence does not change the configuration method on this page.
A knowledge base is made of a library, a collection, and data. The smallest search unit is the library; collections only classify and manage data and do not affect search results. When several knowledge bases are associated for retrieval, they must use the same vector model. The first step in selection Q&A is to check the actual vector model identifier and input type of each library. One original data block can map to multiple vectors; the indexes field in the MongoDB dataset.datas table records the vector ID array. Index content and original content are checked separately. An index written completely and tightly focuses retrieval, while retrieval coverage and index length are weighed against samples. Whether an image can be retrieved depends on whether an image description index or image vector index was created at ingestion time; uploading an image on the query side is not sufficient. Semantic recall and relevance scores affect candidate ordering only and do not carry permission or entity isolation. Whether a retrieval result is usable still requires comparison against the original text.
Constraints at the vector indexes step
The knowledge base search chain is not a single path from user question to vector store to result. Depending on input content and search parameters, FastGPT combines text, image, semantic recall, full-text recall, question optimization, and reranking, then fuses multiple result paths into reference content. Semantic search relies on vector distance and suits natural-language questions and semantically close content. Full-text search relies on keyword hits and suits exact queries such as numbers, models, proper nouns, and error codes. Hybrid search uses both semantic and full-text recall and merges them through RRF. Rerank performs a second ordering on candidate text and suits cases where the text question is clear and candidates are numerous. Image retrieval adds image description or image vector results that fuse with the text side. A result hit by several paths at once typically ranks earlier.
The minimum relevance value is a filter threshold, not a final ordering rule. It takes effect only under semantic search or when result reranking is used. After question optimization, hybrid search, image retrieval, or reranking is enabled, final results may pass through multi-path fusion and do not necessarily follow a single vector similarity ordering. The reference limit caps how many tokens each search may cite. Because chunk lengths differ widely in mixed knowledge bases, a token-based cap is used instead of top k. The dataset search node can be added repeatedly, has external input and static configuration, triggers execution, and outputs references as an array whose length may be zero. A search that returns nothing still lets the downstream output path run.
Configuration and implementation checks
The table below lists documented settings or implementation check items around this page's sample. The numeric values are hypothetical trial starting points for this exercise, used to compare against cited original text from the same question group.
| Configuration or check item | Hypothetical trial value | Comparison method and reason |
|---|---|---|
| Vector model of associated knowledge bases | Identifiers match across libraries; input type is text | Check the model identifier library by library; exclude libraries whose identifiers differ and do not run cross-model retrieval |
| Search mode | Hybrid search | Number-type questions such as commission rates and refund order numbers need keyword hits, while semantic questions need vector recall |
| Result reranking | ReRank enabled | Hybrid search has a wide search range; use the rerank score for second ordering and filtering |
| Minimum relevance | Hypothetical starting point 0.3 | Compare against the rerank score distribution of the same question group and observe whether cited original text falls inside the expected materials |
| Reference limit | Hypothetical starting point 2000 tokens | Compare against the length of a single sample material block to avoid one bill section crowding out reference space |
| Index creation | At least one text index per data block; image blocks built separately per model capability | Check whether the indexes field holds the corresponding vector ID; for image blocks confirm a description or image vector was created at ingestion |
Three common errors
- Associating knowledge bases that use different vector models. The documents state that libraries associated for retrieval must use the same vector model. When identifiers differ, exclude the library and record it. Do not assume a particular recall result will appear after association.
- Uploading an image only on the query side without confirming whether an image index was created at ingestion. Image retrieval depends on a vision model or a vector model that supports image input, and on whether a valid index was generated during ingestion.
- Treating retrieval or form completion as authorization. Retrieving a commission setting record only shows the material can be recalled; it does not show that a given commission was approved. A pending-review status does not mean an operating permission is absent. Signing, deal completion, and contract effectiveness are three distinct facts.
How to verify the setup
Hypothetical exercise sample: the original material text reads "Creator A, March 2025 special session commission rate 15%, settlement bill line 12 corresponds to order 20250310-088, refund deduction detail line 4 cites the same order number, deduction amount 0, status pending review." The draft sentence to be judged is "Creator A's March 2025 special session commission has been settled at 15%."
- Input the draft sentence together with the original text of the three materials above. Compare the settlement status in the draft sentence against the pending-review status in the deduction detail. When both refer to the same order number and the same period and the two values conflict, judge the sentence unsupported by the materials.
- Input the search term "deduction basis for 20250310-088". Compare whether the returned references include the original text of refund deduction detail line 4. When the returned reference matches that line's wording, the condition is met. When the return is empty or points to a different order number, record a miss.
- Input a paper empty result
[]. Check whether the downstream node still runs the reference output path. When the output length is zero and the path continues executing, the condition is met. This result does not mean that any invented rule necessarily returns empty.
This feature has been checked against public technical facts: associated retrieval knowledge bases must use the same vector model, so check the actual model and input type first; one original data block can link to multiple indexes, so check index and original content separately; build indexes with complete and focused information, weighing retrieval coverage and index length against samples; image retrieval also depends on whether an image description or image vector was created at ingestion, not only on the query side; after model or index changes, compare cited original text with the same question group and do not promise results based on vector scores alone.