What the documents look like
For product questions in B2B marketplaces, this page examines FastGPT hybrid retrieval. Compare the material received with the sources used in the answer. The examples are hypothetical.
Constraints at the hybrid retrieval step
A B2B selection question has two distinct concerns: finding a product description and establishing a supplier’s qualification. In the proposed text test, the quotation contains model specifications and the certification material states a business scope. Mark both passages as required candidates. Inspect their presence after RRF fusion of hybrid results, then examine reranking with the complete question. Similar wording between a model description and a business scope is a reading lead; it does not establish actual supply capability. Keep separate entries for the model description found and the qualification evidence still required. A returned quotation must not be logged as a completed supplier review; each finding needs its own source and acceptance step.
Configuration and implementation checks
| Item or implementation check | Input or applicable condition | Traceable check |
|---|---|---|
| Fused candidates | Model description in the quotation; scope in certification material | Retain both originals and check candidate inclusion. |
| Complete question | Compare model and scope wording, without determining supply qualification | Inspect whether the reranking question retains both reading objects. |
| Minimum relevance | Applies to semantic retrieval or reranking | Compare visibility of the scope passage before and after filtering. |
| Citation token limit | Keep model conditions and business-scope wording | Check for an omitted scope restriction beside the quotation name. |
| Missing-source retest | No detailed specification for the customized feature | Repeat after adding the specification and trace the new source. |
Documented conditions
| Documented item / implementation check | Condition |
|---|---|
Hybrid retrieval | Vector and full-text results use RRF fusion. |
| Reranking | Reranking reorders candidates using the full question. |
Minimum relevance | Filtering applies to semantic search or reranking. |
Citation limit | Citation content is limited by tokens. |
Question optimization | Compare subjects and restrictions after optimization. |
Three common errors
In the second proposed example, the quoted product category is absent from the scope stated in the certification document. This requires a comparison between sources. A higher-ranked passage cannot be presented as wording accepted by both documents. Retain the candidate lists before and after sorting, checking whether the scope passage was never recalled or disappeared during later filtering. Even when both are cited, applicability to a procurement decision still requires the relevant party, version and qualification evidence.
How to verify the setup
The customization question has a quotation but no detailed technical specification. Reranking can reorder the supplied quotation text without providing the missing attachment. Record that source gap, add the actual specification when available and repeat the original question. Inspect candidates, filtering, ordering and final citations separately. Check minimum relevance under its semantic-search or reranking conditions; acceptance concerns whether the required specification can be traced, rather than the relevance number alone.
| Material context | Scope |
|---|---|
| Document types | Business Certification Documents, Inquiry and Quotation Sheet, Electronic Contract and Platform Settlement Statement |
| Candidate sources | B2B Platform, National Enterprise Credit Information Publicity System, Payment Transaction Records and Logistics Tracking |
| Technical baseline | FastGPT v4.17.2 |