What the documents look like
A brand concession inside a shopping mall accepts the mall's unified cashiering and commission-based settlement. This single fact shapes the document set. Brand ERP records and mall POS records can diverge in timing, so reconciliation materials such as monthly settlement statements, joint-operation commission contracts, mall activity letters and sales receipts carry the evidence that procurement and supplier questions depend on.
Procurement and supplier materials rarely sit in one clean file. Contracts, quotations and supply lists arrive as attachments and are filed across separate collections. One supplier can map to several documents with different dates and statuses. In FastGPT, the smallest search unit is the library, not the collection. Collections classify data for management, and a search runs across the whole library rather than narrowing to one collection. Questions in this scenario commonly carry a supplier name, a contract number or a material model, which means semantic retrieval and full-text retrieval each have a natural fit.
Constraints at the retrieval step
The knowledge base search node can associate one or more libraries that use the same vector model. Libraries built on different vector models cannot be associated together in one search. The practical step is to check the model identifier and exclude non-matching libraries rather than running a cross-model search to compare what was missed.
Semantic retrieval relies on vector distance. Full-text retrieval relies on keyword hits. Hybrid retrieval uses both and merges the two result sets through RRF, which is a fusion step inside the retrieval chain rather than a standalone switch. The minimum relevance value applies only under semantic retrieval or when result reranking is used. It acts as a filter threshold, and it does not perform permission or entity isolation. The reference cap is measured in tokens and controls how much content one search can cite.
Question optimization performs coreference resolution and question expansion, and it adds an extra model call before retrieval. Search output is an array of references whose length can be zero, and the downstream chain still runs when nothing is found. Retrieval or a form collecting values does not constitute authorization. Approval, database writes and external synchronization require separate downstream implementation.
Image retrieval paths only execute when an image input is present and the model supports the required capability. Reranking only runs when a rerank model is available and enabled by configuration.
Configuration and implementation checks
Hypothetical paper exercise: a concession store manager submits a paper handover request whose draft sentence reads, "Supplier A has confirmed the March 2025 supply list, so procurement can place orders directly." Two sentences from the material under review: first, "Supplier A's quotation is valid until 28 February 2025; the supply period is governed by a separately signed procurement contract"; second, "The March 2025 supply list is pending review and has not been signed." The question is whether the draft sentence's "has confirmed" and "can place orders directly" are supported by the material.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated libraries | Associate only the procurement and supplier library | Check that candidate libraries share the same vector model identifier before associating |
| Search mode | Run semantic retrieval and full-text retrieval separately first | Compare returned content for the sample question before deciding on hybrid retrieval |
| Reference cap | Set to cover the two sample sentences | Test with materials of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Mark relevant and irrelevant samples first, then pick a starting value | Applies only under semantic retrieval or reranking; tune the threshold against the marked samples |
| Question optimization | Retrieve with the original question, then compare with the optimized question | Check whether subject and object are altered, and decide based on observed effect |
| Result reranking | Enable when candidates are numerous | Requires a rerank model enabled by configuration; the rerank score can be used for filtering |
Three common errors
- Reading "pending review" as "has confirmed." In the sample, the supply list is pending review and unsigned. Pending review does not establish that operating permission is missing, and it does not equal a completed signature. Check: input the sample sentences and the draft sentence, compare whether "pending review" and "has confirmed" point to the same field and the same state, and observe whether the draft sentence is supported by the source text.
- Treating quotation validity as the supply period. The sample states the supply period is governed by a separately signed procurement contract, and quotation validity does not indicate the supply period. Check: input the quotation sentence and the draft sentence, compare whether "valid until" and "supply period" are the same field, and observe whether the two are conflated.
- Treating retrieval output as authorization to execute. Returning reference content does not equal approval or permission to place orders. Check: input "can place orders directly" from the request, compare retrieval output against authorization status fields, and observe whether any statement uses retrieved content as an execution basis.
How to verify the setup
- For the draft sentence "Supplier A has confirmed the March 2025 supply list," input that sentence together with the two source sentences. Compare whether "has confirmed" and "pending review, unsigned" refer to the same subject, the same period and the same field, and observe whether the draft sentence lacks support. The expected outcome is that it is unsupported; no live test is claimed.
- For "procurement can place orders directly," input the request and the retrieval output. Compare whether retrieval returns any authorization or approval field, and observe whether a conclusion treats retrieval results as authorization to execute. The expected outcome is that retrieval does not carry authorization, and the actual authorization state remains to be verified.
- For empty-result handling, input a paper input of [] and check whether the downstream design runs the chain with array length zero. Observe whether any claim asserts that a fictional rule must return []. The expected outcome is that the chain runs, with no guarantee of any specific hit.
These checks keep missing evidence missing. A pending request is not an approval, an offered option is not an execution, and an unsigned termination is not a completed termination. Where the material does not state a status, the finding remains unsupported or pending verification rather than being recorded as a denial or a conflict.