What the documents look like
In franchise store supervision, inventory procurement and supplier records are spread across four material types: store visit forms, franchise contracts, rectification notices, and display standards. The supervision task centers on headquarters oversight of how franchisees carry out standards, so a single document can hold headquarters requirements, franchisee commitments, and actual execution records side by side. Materials update on a weekly or monthly basis. One store visit form may record display inspection, headquarters order verification, and rectification tracking together.
For document parsing and chunking, this means a single material contains multiple semantic units: display standard clauses, order quantity records, supplier qualification descriptions, and rectification deadline terms. Chunking needs to preserve the link between a clause and the party it constrains, so that a supplier qualification validity period and a store order quantity do not end up in separate chunks without shared context. Supplier identifiers, contract numbers, and license numbers are exact-match targets; semantic vectors distinguish such strings poorly, and the full-text retrieval path is better suited to carry them.
Constraints at the document chunking step
A FastGPT knowledge base has three layers: library, collection, and data. The smallest search unit is the library; collections serve classification management only. One data item can map to multiple vector indexes, stored in the MongoDB dataset.datas table, where the indexes field records an array of vector IDs. The vectors themselves sit in a PostgresSQL vector field with an HNSW index. Retrieval first recalls vectors, then returns to MongoDB by vector ID for the original data; when the same original data is hit through multiple paths, results merge and the highest score is kept.
Hybrid retrieval fuses semantic and full-text recall through RRF, and ReRank performs a second ordering of candidate text with a 0-1 score. The minimum relevance value takes effect only in semantic retrieval or when result reranking is enabled; it is a filter threshold. The citation limit controls how much content each search cites, measured in tokens. Image retrieval requires that a corresponding index was built at ingestion time and depends on a vision model or a vector model that supports images; reranking requires a ReRank model enabled per settings. These paths are optional, and a single request does not necessarily execute all of them.
Hypothetical paper exercise sample
The following fictional material is used for this paper review only and points to no real store or company.
Material text one: "Appendix three of the franchise contract states that the fresh-food supplier qualification file is valid until 31 March 2026, and the franchisee shall submit renewal materials thirty days before expiry."
Material text two: "The January 2026 store visit form records that the cold cabinet display at a Class A store meets the standard, headquarters order number DD-20260112 has been verified, and the supplier qualification file is archived in volume two of the store records book."
Draft sentence to be judged: "The store's supplier qualification remains valid and renewal materials have been submitted."
Comparison object: same store, same period (January 2026), same field (supplier qualification validity and renewal submission status). The source text states only the expiry date and the deadline for submission; it does not state that renewal materials were submitted, so the second half of the draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | Supplier numbers and natural-language questions coexist, so both query types need coverage; compare number hit behavior against semantic retrieval alone |
| Minimum relevance | 0.3 | Used as this exercise's starting point to avoid filtering out long qualification-validity sentences; compare recall volume against 0.5 |
| Citation limit | 2000 tokens | Carries multiple passages from visit forms and contract appendices; compare against 1000 for whether clause-constrained content is missed |
| Question optimization | Enabled | Follow-up questions such as "what is the second point" need coreference resolution; compare short-question recall against disabled |
| Result reranking | Enabled | Second ordering when candidates are numerous; compare ordering stability against disabled |
| Chunking basis | By clause and record unit | Preserves supplier qualification clauses and the parties they constrain; compare context completeness against fixed-length splitting |
Three common errors
- Reading "renewal materials shall be submitted thirty days before expiry" as "renewal materials have been submitted." The source text states only the submission deadline, not that the submission action was completed. The check should compare whether a submission record exists for the same store and same period; when absent, record it as unsupported.
- Treating a successful upload as retrieval already updated. After a material change, the actual training and index state need checking. Whether the
indexesfield in MongoDB has generated the corresponding vector ID is an observable item; the upload action alone does not show retrieval is in effect. - Treating minimum relevance as a permission or entity filter. This value takes effect only in semantic retrieval or result reranking and filters low-relevance results; it does not exclude unrelated businesses or unauthorized records. Retrieval results and authorization status are separate facts.
How to verify the setup
- Input material text one and material text two, and compare whether the draft sentence's "renewal materials have been submitted" has a corresponding record. If neither source text contains a submission status field, the sentence is unsupported and cannot be judged as submitted or rejected.
- Input a supplier qualification number or contract number, and compare hit differences between full-text retrieval and semantic retrieval. If a number-type query hits under full-text retrieval and misses under semantic retrieval, the chunking's retention of exact strings needs inspection, and the actual observation is recorded.
- After a material update, compare the upload time with the actual state of the
indexesfield indataset.datas. Ifindexeshas not generated the corresponding vector ID, retrieval will not return the new content; record this as index not updated, without inferring training failure or data loss.