What the documents look like
Promotion registration work is carried by four document types: activity registration forms, promotional price approval sheets, inventory lock tables, and post-campaign review tables. Together they carry price, inventory, and campaign rhythm. Advertising copy review sits outside this page's task.
A single item can appear with several price fields at once. Registration price, campaign price, and lowest final price are separate fields. A contradiction exists only when one field holds mutually exclusive values in the same period. The inventory lock table records locked quantity and lock scope. The review table records actual results after the campaign closes. Materials update weekly or per campaign. After an update, the actual training and index state must be checked; a successful upload does not mean retrieval has refreshed. People in this work include platform campaign operators, brand sales planners, and finance business partners. Retrieved candidates serve question answering only. They do not grant registration, approval, or inventory-lock authorization.
Constraints at the document chunking step
Price fields and inventory fields often appear as tables, conditional statements, and footnotes. When chunking separates a condition from the price it governs, the retrieved fragment loses its effective scope.
A FastGPT knowledge base has three parts: library, collection, and data. The smallest search unit is the library. A collection only classifies data and does not affect search results. Vectors are stored in a PostgresSQL vector field with an HNSW index. The MongoDB dataset.datas table stores source data and provides an indexes field that records vector IDs; one data group can map to multiple vectors. At retrieval, vectors are recalled first, then the vector ID is used to find source data in MongoDB. Records matching the same source data are merged, and the vector score keeps the highest value.
Semantic search relies on vector distance. Full-text search relies on keyword hits. Hybrid search merges both paths through RRF. ReRank runs a second ordering over candidates and returns a 0-1 score. The minimum relevance value is a filter threshold only, effective under semantic search or when result reranking is enabled. It does not perform permission or entity isolation. A retrieval library and its vector model must be associated from the same source. Libraries bound to different vector models cannot be associated together. In that case, check the identifiers, exclude the mismatched library, and record the actual observation without inferring runtime results.
Configuration and implementation checks
The table below covers this page's sample. Values are hypothetical starting points for this trial, used to compare against sample retrieval behavior. They are not official defaults.
| Configuration item | Hypothetical trial starting point | Reason and comparison object |
|---|---|---|
| Search mode | Hybrid search | The sample mixes coded fields and natural-language questions, so semantic recall and full-text hits are both examined; compare against semantic search alone |
| Minimum relevance | 0.3 | Starting point to observe whether low-score fragments are filtered; compare returned count and content range against 0.5 |
| Citation limit | Measured in tokens, starting point 2000 | Chunk lengths vary widely in a mixed library, so tokens control citation volume; compare citation completeness against 1000 |
| Question optimization | Off | The sample question has a clear subject, avoiding an extra model call; compare latency and recall against the enabled state |
| Result reranking | Enabled | With many candidates, ReRank scores reorder them; compare ranking position against the disabled state |
| Index check | Verify the indexes field against source data correspondence | Confirm whether a valid index was generated at ingestion; compare against checking upload status alone |
Three common errors
- Treating registration price, campaign price, and lowest final price as one field in conflict. Check: input the sample text, compare the three field names with their respective values, and observe whether one field holds mutually exclusive values; when the fields differ, record the result as unsupported.
- Treating a successful upload as refreshed retrieval. Check: input the updated material and one term query, compare the returned content against the updated text, and observe whether the return still shows the old version; when training or indexing has not updated, record the gap.
- Treating semantic search or minimum relevance as a permission filter. Check: input a query term touching an unauthorized scope, compare returned candidates against the access list, and observe whether the return includes records outside access rights; retrieval does not perform authorization decisions and needs a separate system implementation.
How to verify the setup
The following verification uses a hypothetical drill sample. The sample text is: "Activity number A-2025-0317, product A, registration price 199 yuan, campaign price 189 yuan, lowest final price 179 yuan, price effective range 2025-03-17 to 2025-03-20. Inventory lock quantity 500 units, lock scope East China warehouse only." The draft sentence to be judged is: "Product A's lowest final price for this campaign is 179 yuan, valid until 2025-03-20." The subject, object, and field in this sentence come from the source text, so it is supported within the source text, but the validity wording narrows the price effective range to a single end date and must be compared against the source start and end range, with the difference recorded.
- Input the sample text and the draft sentence, compare the price field name and value in the draft against the matching field in the source, and observe whether they share the same subject, same period, and same field; a matching field with an identical value meets the condition, while a replaced field or rewritten period records as unsupported.
- Input a paper empty result
[]and a downstream citation assembly design, compare the output after the empty result enters the citation step, and observe whether downstream handles it as an empty set; an empty output or an explicit no-citation notice meets the condition, while other output records a gap. - Input a cross-library identifier list, compare the vector model identifier bound to each library, and observe whether a same-source association exists; matching identifiers record as associable, and mismatched identifiers are excluded with the actual observation recorded, without running a cross-model retrieval to compare missed recalls.
This page covers only: platform promotion registration and price management / product material and selection Q&A / document parsing and chunking. During operations handover, the authorization state of materials, operations, and requests must be clear; retrieval or a form does not execute authorization.