What the documents look like
Parcel handoff happens at the point where a package leaves the warehouse, and the central concern is carrier scan responsibility. The supplied brief lists the daily documents for this category as electronic waybills, handoff manifests, pickup scan records, and exception item registrations, with updates occurring per batch and daily. This page addresses only knowledge base retrieval and recall inside product information and selection questions. The retrieval target is the knowledge base content formed from these documents. Retrieval results do not carry authorization judgments.
A knowledge base is organized into library, collection, and data layers. The smallest search unit is the library, and search runs against the whole library; collections serve classification management and do not affect search results. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Libraries using different vector models cannot be associated together. The correct action is to check the vector model identifier, exclude non-matching libraries, and record the actual observation. Cross-model retrieval is not performed to compare missed recalls.
Retrieval input can include text and images. An image participates in retrieval only when a corresponding index was built at ingestion time and a vision model or an image-capable vector model is configured. The documented overall flow lists semantic retrieval, full-text retrieval, hybrid retrieval, question optimization, rerank, and image retrieval as optional paths; not every request executes all of them. Rerank requires a rerank model and must be enabled per settings. Reference content is output as an array whose length can be zero, and the output chain still runs when nothing is found.
Constraints at the retrieval step
The search mode choice affects which candidates are recalled. Semantic retrieval relies on vector distance and suits natural language questions and semantically similar content. Full-text retrieval relies on keyword hits and suits identifiers, model numbers, proper nouns, and error codes. Hybrid retrieval uses both and merges results through RRF. RRF is a fusion process for multiple result paths in hybrid retrieval, not a configuration switch that can be turned on. A rerank model produces a 0-1 score representing relevance between the searched content and the question, and that score can be used for filtering. The minimum relevance value is a 0-1 filter threshold and takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold, not a final sorting rule, and it does not exclude unrelated businesses or unauthorized records. Question optimization performs coreference resolution and question expansion before retrieval; it adds a model call and therefore increases overall latency.
Reference upper limit is expressed in tokens per search, not as a top-k item count. The documented reason is that chunk lengths vary widely in mixed knowledge bases, which makes top-k results unstable. When a rerank model is available and enabled, FastGPT merges reranked results, vector search results, and full-text retrieval results through RRF to produce the final search result. The final reference content is not necessarily ordered strictly by a single vector similarity score; content hit by multiple recall paths is usually ranked higher.
A hypothetical paper exercise sample follows. Source text one: "Electronic waybill number range for this batch is A-0001 to A-0050. The handoff manifest states delivery to carrier branch Jia, handoff date 2026-03-10. Pickup scan record status shows 48 scanned and 2 unscanned." Source text two: "Exception item registration states waybill A-0037 has a damaged label, registration date 2026-03-10, handling status pending verification." Draft sentence to be judged: "All 50 items in this batch have completed carrier pickup scanning." The draft sentence and the pickup scan record share the same subject, the same period, and the same field. The record shows 2 unscanned items, and the same-field values are mutually exclusive. The judgment is contradiction.
Configuration and implementation checks
| Configuration or check item | Value or action in this page's sample | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases using the same vector model | The node can associate one or more knowledge bases with the same vector model |
| Search mode | Compare the sample draft sentence under semantic, full-text, and hybrid retrieval | No single mode is presumed suitable for all material |
| Reference upper limit | Hypothetical trial starting point of 2000 tokens, compared with 800 tokens for content completeness | Unit is token; the starting point is set for this exercise and is not an official default |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust the value | Takes effect only under semantic retrieval or result reranking |
| Question optimization | Compare the original question with the optimized question and check whether the subject was altered | Used for coreference resolution and question expansion |
| Empty result handling | Use paper input [] to check the downstream design | Reference content length can be zero and the output chain still runs |
Three common errors
- Treating retrieval as authorization. A retrieval hit on a pickup scan record does not mean the carrier has completed responsibility confirmation, and it does not mean an exception item has been handled. The input is the search term and library content; the comparison is textual relevance; the observable result is reference content. Authorization status requires a separate system implementation.
- Treating minimum relevance as a permission or entity filter. The value is a 0-1 filter threshold and does not exclude unrelated businesses or unauthorized records. The input is the question and candidate content; the comparison is similarity or rerank score; the observable result is a change in the number of reference items.
- Describing RRF as a switch. RRF is the fusion process for multiple result paths in hybrid retrieval, not an enableable configuration item. The input is the semantic and full-text result paths; the comparison is the ordering before and after fusion; the observable result is that a single content item hit by multiple paths ranks higher.
How to verify the setup
- Input source text one and the draft sentence "All 50 items in this batch have completed carrier pickup scanning." Compare the pickup scan record's 48 scanned and 2 unscanned against the draft sentence's all-items statement. The condition is met when the retrieval reference contains the record text and the draft sentence is marked as contradicting the record. The expected result is judged on this basis and is not claimed to have been tested.
- Input source text two and the query "A-0037 damaged label handling status." Compare the exception item registration's pending verification status against the query intent. The condition is met when the reference content contains that registration and does not include a conclusion such as "handling completed" that the record does not state. The failure branch is an empty reference; in that case the gap is recorded and no cause is inferred.
- Input the paper empty result [] and the downstream handling design. Compare whether downstream continues execution with a reference of length zero. The condition is met when the output chain runs and downstream does not produce fabricated references. No claim is made that a hypothetical rule retrieval necessarily returns [].
Libraries using different vector models cannot be associated together; check identifiers first and exclude non-matching ones. Pending review does not demonstrate a lack of operating permission. A quotation validity period does not indicate a supply period. Anything not stated, or supported only by related-stage evidence, is unsupported or pending verification.