What the documents look like
Site-selection work at a chain headquarters happens before a store opens. The core materials are the site evaluation form, the letter of intent for lease, the renovation budget, and the opening schedule. These are maintained by headquarters expansion staff, regional development staff, and engineering managers, and they are updated on an irregular basis, once per store. The procurement and supplier records attached to this stage carry similar characteristics: supplier profiles, historical procurement scopes for comparable store types, quotation validity periods, and delivery milestones.
In a vector index, these materials are split into data chunks. Each chunk may be mapped to more than one index, so a single original data block can be recalled through several index entries. The smallest search unit in FastGPT is the knowledge base, not the collection; collections only classify data and do not affect search results. The vector store uses PostgresSQL with the PG Vector plugin and an HNSW index, while MongoDB stores the original data in the dataset.datas table, including an indexes field that records the vector IDs belonging to that data. During retrieval, vectors are recalled first, then the vector IDs are used to locate the original content in MongoDB; entries that map to the same original data are merged and the highest vector score is kept.
Constraints at the vector indexes step
Knowledge bases linked together for retrieval must use the same vector model. Before linking, check the actual model identifier and the input type each model accepts. Knowledge bases with different vector models cannot be linked together for retrieval; exclude the mismatched base, record the observation, and do not run a cross-model retrieval to compare recall.
Index content should be complete and focused. Retrieval coverage and index length need to be weighed against the sample at hand. An original chunk may carry multiple indexes, and the index content and the original text should be checked separately.
Image retrieval depends on whether an image description index or an image vector index was generated at ingestion time, not only on what is submitted at query time. Full-text search suits exact lookups such as numbers, model codes, and proper nouns. Hybrid search merges multiple recall paths through RRF, and result reranking performs a second ordering of candidate text. The reference limit is controlled by tokens. The minimum relevance threshold takes effect only in semantic retrieval or when result reranking is used.
A hypothetical paper exercise sample: an expansion staff member at a chain headquarters intends to check a draft site evaluation form. The source text contains two sentences: "Object: an intended store in the East China region; Status: the letter of intent for lease has been signed at this stage; Scope: the renovation budget is scheduled backward from the opening schedule." The draft sentence to be judged is: "The lease for this store has taken effect, and the supplier may deliver according to the opening schedule." The check point is that the source states only that the letter of intent was signed. It does not state that the lease contract has taken effect, and it does not state any supplier delivery commitment. The draft sentence is unsupported.
Configuration and implementation checks
The table below lists only settings or implementation checks that the official documents describe and that relate to this page. The hypothetical trial starting values are for this paper comparison only; the reason for each choice and the comparison object are stated alongside.
| Configuration or check item | Hypothetical trial starting value | Reason and comparison object |
|---|---|---|
| Vector model of linked knowledge bases | Same vector model | Check each base's model identifier first, exclude differing models, and do not run cross-model retrieval |
| Search mode | Hybrid search | Compare against semantic search and full-text search on the same set of questions |
| Reference limit | A starting value set in tokens | Compare the length of quoted source text at different starting values |
| Minimum relevance | A starting value between 0 and 1 | Takes effect only in semantic retrieval or result reranking; compare results before and after filtering |
| Question optimization | Decide whether to enable based on actual results | Compare retrieval results for the same question with the feature off |
| Result reranking | Available model, enabled according to settings | Compare candidate order before reranking and observe score changes |
Three common errors
- Linking knowledge bases that use different vector models. The input is the model identifier of each base; the comparison is whether the identifiers match; the observable result is that a mismatched identifier requires exclusion and retrieval does not proceed.
- Treating the signing of a letter of intent as the lease taking effect. The input is the status field in the sample source text; the comparison is whether the source states that the contract has taken effect; the observable result is that when the source states only that the letter of intent was signed, the draft sentence "the lease has taken effect" is unsupported.
- Treating a similarity score as a permission or entity filter. The input is the score returned by retrieval; the comparison is the score against the authorization status field; the observable result is that the score does not change whether a record falls within the currently accessible range.
How to verify the setup
- Input the sample source text and the draft sentence. Compare the status field in the source with the wording of the draft. If the source states only that the letter of intent was signed, the draft sentence "the lease has taken effect" is judged unsupported, and the evidence gap is retained.
- Input the same set of questions. Compare the vector model identifiers of the linked knowledge bases. If a differing model identifier is present, exclude that base and record it. Do not run a cross-model retrieval to compare missed recall.
- Input a clearly defined paper empty result. Compare the downstream quoted-content output. If the output is an array of length 0, the empty-result path is confirmed to pass through. Do not claim that a fictional rule's retrieval necessarily returns an empty array.