What the documents look like
Convenience-store fresh food is ordered by the store at high frequency against short shelf life, and it is not handled as a joint-operation counter with revenue sharing. The working documents for this category are the fresh-food order form, the damage write-off form, the near-expiry handling sheet, and the temperature record. These are produced per day or per shift, drawing on POS sales, the headquarters ordering system, store inventory, and cold-chain temperature records.
Operating reports and exception reviews depend on being able to trace each statement back to a specific document, a specific field, and a specific version. The order form carries the store, the item, the ordered quantity, and the ordering-system status. The damage write-off form and the near-expiry handling sheet carry their own recorded states. The temperature record carries time-stamped readings. A report sentence that describes approval, completion, or write-off must be checked against the field that would carry that state in the named source document.
Two facts shape this task. First, order forms move through states such as pending review, and a pending state is not an approval. Second, the source material is refreshed daily or per shift, so a citation must be checked against the current version and any prior version, and use of the citation still depends on valid-version review. The retrieval step in FastGPT outputs a citation array whose length may be zero; an empty array means no content was found, and that output path still runs, so downstream design has to accept it.
Constraints at the citation tracing step
Knowledge-base search first verifies the selected knowledge bases and the material version. Only knowledge bases using the same vector model may be linked together. Results are passed to the answering step, and each answer sentence is checked against the cited original text. A thinking process or a fluent answer does not replace source evidence.
The minimum relevance value is a filter threshold. It does not perform permission or entity isolation, and retrieval is not authorization. Mixed retrieval merges multiple result paths through RRF. Reranking requires a rerank model and must be enabled in the settings. Image retrieval requires image input and model capability. The overall retrieval flow lists optional paths; not every request executes all of them.
A pending request is not an approval, an offered option is not execution, and an unsigned termination is not a completed termination. When a source field does not state a state, that state remains unverified. A judgment requires the same subject, the same period, and the same field; only a clearly mutually exclusive value in that same field counts as a contradiction.
Configuration and implementation checks
The table below uses a hypothetical drill sample and lists official documented settings or implementation check items relevant to this page's main task. The values are starting points for this hypothetical trial, not official defaults.
| Configuration or check item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | One knowledge base per vector model, all sharing the same vector model | Check whether the vector model identifiers match; exclude mismatched ones and do not run a cross-model retrieval comparison |
| Search mode | Mixed retrieval | Compare against semantic retrieval alone and observe whether numbered queries are hit |
| Minimum relevance | Set to 0.3 | Compare against 0.5 and observe the difference between citation count and manual verification results |
| Citation limit | Set to 2000 tokens | Compare against 1000 tokens and observe whether key fields are truncated |
| Result reranking | Enabled | Compare against disabled and observe whether candidate ordering better matches the question subject |
| Question optimization | On | Compare against off and observe whether short questions have their subject completed |
Three common errors
- Treating retrieval output as an authorization conclusion. Observable error: the report states that a write-off was approved, but the citation comes only from the retrieval array and no approval field is present. Check: input a retrieval result containing the original write-off form text, compare whether the original text states an approval status, and observe whether a missing status is marked as pending verification.
- Treating an empty citation array as process termination. Observable error: downstream design does not accept an empty array, so the report is missing a section. Check: input a paper-level
[], compare whether downstream outputs "no citation retrieved," and observe whether execution continues. - Forcing knowledge bases with different vector models to be linked. Observable error: two knowledge bases with different vector models are linked and a missed recall is claimed. Check: input two knowledge bases with different vector model identifiers, compare whether the identifiers match, and observe whether mismatched ones are excluded and the actual observation is recorded.
How to verify the setup
Hypothetical drill sample: the source text reads, "On March 10, 2025, Store A submitted a fresh-food order form for 30 units, and the ordering system recorded the status as pending review. On the same day, the temperature record showed the refrigerated cabinet at 5°C." The draft sentence to be judged reads, "The order form has been approved and the write-off completed." Judgment: the source text states only pending review and does not state approval or completed write-off. After comparing the draft sentence against the source text for the same subject, same period, and same field, the draft sentence is unsupported.
- Verification one: input the source text above and the draft sentence, compare whether "pending review" and "has been approved" are mutually exclusive values in the same field, and observe whether the draft sentence is marked as unsupported.
- Verification two: input a paper-level
[]as the retrieval output, compare whether the downstream design accepts an empty array, and observe whether it outputs that no citation was retrieved while the flow continues. - Verification three: input two knowledge bases with different vector model identifiers, compare whether the identifiers match, and observe whether mismatched ones are excluded and the gap recorded, without running a cross-model retrieval comparison.
These checks describe what to compare and what to observe. They do not assert that a test, deployment, approval, or refund has occurred. Where a source field is absent, the evidence gap remains open.