What the documents look like
Chain-store site selection and store opening work happens before a store begins trading. The core of the work is lease negotiation and construction handover. When this work goes wrong, the consequences include an unprofitable store, delayed delivery, and loss of a deposit. The documents handled day to day include site evaluation forms, letters of intent for lease, renovation budgets, and opening schedules. The data sources behind them include trade-area footfall, map points of interest, historical store-type data from POS records, and landlord materials. Updates are irregular and occur per store. This page does not infer batch synchronization or real-time mechanisms from that update pattern.
Product information and selection questions have to work with the product details, model numbers, identifiers, and natural-language descriptions inside these documents. Document parsing and chunking determine the granularity at which that content enters the retrieval library, and therefore directly affect whether later questions can locate the corresponding conditions and scope.
In FastGPT, a knowledge base consists of libraries, collections, and data. The smallest search unit is the library; collections exist for classification and management. One group of data can correspond to multiple vectors, and the indexes field records the corresponding vector IDs. During retrieval, vectors are recalled first, then the vector IDs are used to find the original data in the dataset.datas table in MongoDB. When the same group of original data is matched, the results are merged and the vector score takes the highest score. Chunking therefore needs to preserve the document identity and version that allow a return to the source text, and it needs to be checked for whether conditions, explanations, and the content they constrain have been split apart. Semantic retrieval depends on vector distance, full-text retrieval depends on keyword hits, hybrid retrieval merges multiple result paths through RRF, and Rerank performs a second ordering. A similarity setting is a filter threshold; it does not carry business permissions or entity isolation.
Hypothetical paper exercise sample: a chain headquarters development department intends to check a "East China Region Site Evaluation Form (No. XZ-2026-014)" against a "Letter of Intent for Lease (No. ZY-2026-031)". The evaluation form states: "The store type covered by this form is the community store; the evaluation period is 1 March 2026 to 31 March 2026; the upper limit for the rent-to-sales ratio calculation is 12%." The letter of intent states: "This letter of intent takes effect from the date both parties sign it; the lease terms are subject to the formal contract; the intent deposit is not counted toward rent." The draft sentence to be judged is: "XZ-2026-014 has confirmed that the rent-to-sales ratio at this site does not exceed 12%, and ZY-2026-031 has been signed and has taken effect."
Constraints at the document chunking step
The evaluation form, the letter of intent, the renovation budget, and the opening schedule each carry conditions that belong to a specific store type, period, and scope. A chunk that keeps the subject, the object, and the numeric range together remains usable for selection questions. A chunk that separates the store type from the rent-to-sales ratio limit, or that drops the document number, loses the ability to point back to the source.
The multi-vector design matters here. A longer passage can be mapped to several vectors, so that a hit on any one of them recalls the same group of original data. This supports keeping a complete condition in one data group while still indexing parts of it. The merge rule then applies: when the same original data is hit, the vector score takes the highest score.
The retrieval mode also shapes what chunking must preserve. Semantic retrieval suits natural-language questions and semantically close content. Full-text retrieval suits identifiers, model numbers, proper nouns, and exact lookups, which is relevant when a question names XZ-2026-014 or ZY-2026-031. Hybrid retrieval uses both and merges through RRF. Rerank performs a second ordering and produces a 0-1 score that can be used for filtering. The minimum relevance value applies only in semantic retrieval or when result reranking is used, and it is a filter threshold, not a final ordering rule.
Configuration and implementation checks
| Configuration or check item | Content for this page | Implementation check method |
|---|---|---|
| Chunk granularity | Site evaluation form, letter of intent for lease, renovation budget, opening schedule | Check whether each passage retains the document number, the applicable store type, and the period |
| Multi-vector mapping | One group of data can correspond to multiple indexes | Check whether long text has multiple index groups and confirm the merge rule for original data |
| Search mode | Semantic retrieval, full-text retrieval, hybrid retrieval | Search for the identifier XZ-2026-014 and for a natural-language phrasing separately, and record the results |
| Reference limit | Each search cites at most n tokens of content | For this hypothetical exercise, the starting value is set at 3000 tokens, because it covers the full text of the two samples; compare it with the actual token volume returned by retrieval |
| Minimum relevance | A 0-1 value, effective only in semantic retrieval or with result reranking | Record the threshold and the returned items, and confirm it is not treated as a permission filter |
| Result reranking | ReRank requires a model and must be enabled by configuration | Check whether reranking is enabled and what rerank score is produced; when it is not enabled, record that this path did not run |
Three common errors
- Treating "takes effect from the date both parties sign it" in the letter of intent as a completed signing that has already taken effect. The source text states only the condition for taking effect; it does not record the fact of signing. The phrase "has been signed and has taken effect" in the draft sentence is not supported.
- Treating "the upper limit for the rent-to-sales ratio calculation is 12%" in the evaluation form as confirmation that the site's rent-to-sales ratio does not exceed 12%. The source text gives a calculation upper limit, which is an evaluation basis. The phrase "has confirmed that the rent-to-sales ratio at this site does not exceed 12%" goes beyond the source scope.
- Treating a successful upload or a submitted form as proof that retrieval has been updated. The documentation states that data must be confirmed as trained. A successful upload and the index state are two different things.
How to verify the setup
- Input the source text of the "East China Region Site Evaluation Form (No. XZ-2026-014)" from the sample and search for "community store rent-to-sales ratio upper limit". Compare whether the returned content retains the complete conditions "the store type covered is the community store" and "the upper limit for the calculation is 12%". The condition is met when the returned passage contains the subject, the object, and the numeric range together and can point back to No. XZ-2026-014. It is not met when conditions are split apart or the number is lost.
- Input the draft sentence "XZ-2026-014 has confirmed that the rent-to-sales ratio at this site does not exceed 12%, and ZY-2026-031 has been signed and has taken effect" and check it item by item against the two sample source texts. Compare whether "has confirmed" and "has been signed and has taken effect" have source support. The source texts give only a calculation upper limit and a condition for taking effect, so the judgment is that these phrases are unsupported. If a returned result claims confirmation or effect, record it as going beyond the source text.
- Input "Is the intent deposit in ZY-2026-031 counted toward rent?" and search the source text of the letter of intent. Compare whether the return contains "the intent deposit is not counted toward rent". The condition is met when the sentence is present and can point back to No. ZY-2026-031. When an empty result is returned, record that retrieval did not hit, keep the gap open, and do not infer that the letter of intent lacks this clause.