What the documents look like
Peak-sale warehousing and delivery scheduling centers on short-term peak capacity and the arrangement of people, vehicles and space, which distinguishes it from steady multi-warehouse inventory synchronization. The main data sources are historical peak-sale orders, platform campaign registrations, WMS capacity and courier pickup capability. The update rhythm follows each campaign and each day. For product and marketing copy drafting, the retrievable material sits in peak-sale forecast tables, staffing rosters, courier assurance plans and exception daily reports.
These documents carry explicit time ranges and responsible parties. A commitment sentence in drafted copy has to match the source material on the same responsible party, the same period and the same field. A roster that states a per-shift ceiling for temporary sorting staff at a named warehouse during a named date range supports only that ceiling, that warehouse and that range. A courier assurance plan that states a pickup window supports only that window. An exception daily report that records a status supports only that status as recorded.
Constraints at the retrieval step
In FastGPT, a knowledge base consists of libraries, collections and data. Collections classify data for management, and the minimum search unit is the library, so a search runs against the whole library. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together, so the vector model identifier is checked first and non-matching libraries are excluded rather than searched across models.
Retrieval paths include semantic search, full-text search, hybrid search, image description retrieval, image vector retrieval and result reranking. RRF fuses multiple result channels after hybrid search. The reference cap is measured in tokens. The minimum relevance value is a 0-1 number that applies only under semantic search or when result reranking is used, and it is a filter threshold rather than a final sort rule. When question optimization, hybrid search, image retrieval or reranking is active, the final result may come from multi-channel fusion and need not follow a single vector similarity ordering. Image retrieval requires an image input and the corresponding model capability; reranking requires a rerank model and being enabled in settings. Retrieval does not perform permission or entity isolation, and a similarity score does not separate business ownership or authorization.
Hypothetical paper exercise sample. Source text for comparison: "During the 2025 Double Eleven peak sale, the per-shift ceiling for temporary sorting staff at the East China warehouse is 120 people, applicable from November 1 to November 3." Draft sentence to be assessed: "During the peak sale, the East China warehouse can reach 150 sorting staff per shift." Comparing responsible party (East China warehouse), period (November 1 to 3) and field (per-shift headcount ceiling), the draft value exceeds the source range and is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only libraries that contain peak-sale scheduling documents and share one vector model identifier | Check the vector model identifier first, exclude non-matching libraries, and do not run cross-model retrieval |
| Search mode | Hybrid search | Compare recalled content against semantic search and full-text search on the same test question |
| Reference cap | Hypothetical starting point of 2000 tokens | Test whether long and short rosters and assurance plans return complete content, then adjust |
| Minimum relevance | Hypothetical starting point of 0.4 | Label relevant and irrelevant samples first, then compare what survives filtering |
| Question optimization | Enabled | Compare the original question with the optimized question and check whether the responsible party was altered |
| Result reranking | Enabled per settings with a model available | Compare ordering against results without reranking, and use the rerank score for filtering |
Three common errors
- Treating retrieval as permission filtering. The minimum relevance value is a similarity threshold and does not exclude unrelated business records or unauthorized records. The check is to confirm that the retrieval scope is already limited to libraries the operator may access, and to record the evidence gap when that confirmation is missing.
- Treating a pending-review status as approved. A "pending review" entry in the material does not demonstrate the absence of operating permission and does not demonstrate that authorization has been granted. The check is to compare the status field in the source text with the wording of the draft sentence, and to keep the item as pending verification when the status does not match.
- Treating a quotation validity period as a supply period. These are different fields, and different values within the same period do not constitute a contradiction. The check is to compare field by field, and to record an unsupported item when no supply period is stated.
How to verify the setup
- Input the hypothetical source text above together with the draft sentence "During the peak sale, the East China warehouse can reach 150 sorting staff per shift," and compare responsible party, period and field. The source ceiling is 120 people and the draft value exceeds it, so the observed result is unsupported.
- Input a continuous-conversation follow-up that contains only "What is the second point," with question optimization enabled, and compare the optimized question with the original. The optimized question should restore the referenced subject; if the subject is altered, record the deviation.
- Input an explicit paper-level empty result [] and check whether the downstream design carries the empty-reference path. The reference output can have a length of zero, and no claim is made that a fabricated-rule search necessarily returns [].
A relevance score is not a category, entity or permission filter, and retrieval is not authorization. Forms collect values; business verification, approval, database writes and external synchronization require their stated downstream implementation.