What the documents look like
Franchise filing management handles legal disclosure and the commerce department system. It does not handle individual site engineering. The document set for this page is the franchise contract, the operations manual, the filing materials, and the information disclosure file. Sources named in the brief are the commercial franchise information management system, enterprise licenses, store lists, and the contract ledger.
These materials carry subject, object, period, and status fields. A contract has parties and a termination clause. A disclosure file has a filing year and a version. A store list has a scope of outlets. A filing material set has a submission status. Updates are irregular or annual. Upload success and completed training indexing are two separate states. This page covers only the parsing and chunking step.
The category differs from headquarters site selection and store opening because the work product is a legal disclosure package submitted to a government system, not a floor plan or a lease. A chunk that separates a condition from the clause it governs will retrieve a fragment without its scope of application. A chunk that drops the document identity or version cannot be traced back to the source text it came from.
Constraints at the document chunking step
Franchise filing documents are mostly clause-structured. One paragraph often carries an applicable subject, an applicable store scope, and an effective condition at the same time. Chunking must keep the document identity and version that allow a return to the original text. Paragraph splitting must not separate a condition from its conclusion.
In FastGPT, a knowledge base is made of library, collection, and data. The smallest search unit is the library. A collection is for classifying data and does not affect search results. One group of data can map to multiple vectors. The indexes field records vector IDs, so several vectors can point back to the same group of original data. When retrieval hits several of those vectors, they merge into the same original data group and the vector score takes the highest value.
Retrieval recalls similar content. It does not perform permission or entity isolation, and it does not decide whether a filing has been approved. A similarity score is not a category, entity, or permission filter. A retrieval hit is not authorization.
Hypothetical paper test case: The source material reads, "Party A is the franchisor headquarters, Party B is the franchisee. This disclosure file applies to stores newly signed in the 2024 filing year. Contract termination requires written signature by both parties." The draft sentence to be judged reads, "This disclosure file covers all franchise stores." The comparison input is the three source sentences above together with this one draft sentence. The comparison object is the source wording "stores newly signed in the 2024 filing year" against the draft wording "all franchise stores." The source text does not state all stores, so the draft sentence is unsupported.
Configuration and implementation checks
The table below is a hypothetical starting point for this paper exercise. The values are for comparison observation only and are not official defaults.
| Configuration item | Hypothetical starting value | Reason and comparison object |
|---|---|---|
| Search mode | Hybrid search | The sample contains a year-type identifier, so semantic and keyword recall are both observed |
| Reference upper limit | 2000 tokens | The document describes control by tokens; compared against the actual length of the sample paragraph |
| Minimum relevance | 0.3 | Takes effect only in semantic search or with reranking; compared against unfiltered results |
| Question optimization | Enabled | The sample question contains a reference; the rewritten query is observed |
| Result reranking | Enabled | Used for secondary ordering when candidates are many; the rerank score is observed |
| Chunking method | By clause boundary | Compared against fixed-length splitting for completeness of the clause |
The reference upper limit is expressed in tokens, not in a top-k item count. Minimum relevance is a filter threshold, not a final ordering rule. RRF is a fusion process for combining multiple recall paths; it is not a toggle. Reranking requires a model and must be enabled by a setting. Image retrieval requires an image input and a model capability, and is not executed on every request.
Three common errors
- Treating upload success as completed indexing. Input is a changed disclosure file. Compare the upload state against the training state. Observe whether retrieval still returns content from the earlier version. The document states that upload success and completed training indexing are two states, so an upload record alone does not show that the index reflects the new text.
- Treating a retrieval hit as authorization. Input is a franchisee identity question. Compare the returned fragment against the authorization record. Observe whether similar text is read as an approved authorization. Retrieval recalls similar content and does not perform permission or entity isolation, so a returned fragment does not establish an authorization fact.
- Treating minimum relevance as an entity filter. Input is a question about another brand. Compare the returned results against the target subject. Observe whether similar content is taken as material of this subject. Minimum relevance is a similarity threshold that takes effect only in semantic search or with reranking, and it does not exclude unrelated businesses or unauthorized records.
How to verify the setup
- Check the scope wording against the hypothetical source text and the draft sentence. Input the three source sentences and the draft sentence. Compare "stores newly signed in the 2024 filing year" with "all franchise stores." The condition is met when the draft sentence is marked unsupported, because the source does not state all stores.
- Run a full-text search on the identifier in the sample. Input "2024 filing year." Compare whether the returned fragment contains the complete clause in which the identifier appears. The condition is met when the fragment keeps the condition sentence attached to the clause it governs.
- Check downstream handling of an empty result. Input the paper input
[]. Compare whether the downstream design treats it as an empty result. The condition is met when no fabricated rule content is returned. This is a paper check of downstream design, not a claim that a retrieval of fictional rules returns an empty set.
For each check, record what was compared and what remains unknown. Where a document field is not provided, the field stays unknown. A pending request is not an approval. An offered option is not an execution. An unsigned termination is not a completed termination. Signed, cooperation reached, and contract effective are three different facts. Where the source text does not state an authorization or effective clause, that evidence gap is preserved rather than filled.