What the documents look like
Livestream after-sales work concentrates after a single broadcast session, and the materials that drive it are strongly tied to what the host promised on air. The four document types in scope are livestream order lists, refund applications, reshipment orders, and script templates. Selection questions that reach a knowledge base usually ask about a specific product, model, batch, or commitment wording, while the retrieval terms themselves are short and mixed, carrying order numbers, refund numbers, and reshipment reasons.
Material updates follow a per-session and daily rhythm. That rhythm affects the text range the index has to cover. It does not imply batch synchronization, transaction-triggered automation, or any other mechanism that the documents do not describe.
Constraints at the vector indexes step
A FastGPT knowledge base is built from libraries, collections, and data. The smallest search unit is the library; collections only classify data and do not change search behavior. Vector storage uses the PostgresSQL PG Vector extension with an HNSW index. The MongoDB dataset.datas table stores the original data, and its indexes field records an array of vector IDs, so one data item can map to several vectors. During retrieval, vectors are recalled first, then the vector IDs are used to find the original data in MongoDB. Items that resolve to the same original data are merged, and the highest score is kept.
Knowledge bases linked to the same search node must share the same vector model. This constrains the node input selection only; it does not extend to external systems. Hybrid search merges semantic and full-text results through RRF, which is a fusion process and has no corresponding on/off switch. Reranking requires a ReRank model enabled by configuration, and image retrieval requires image input plus the matching model capability. These optional paths do not all run on every request. The minimum relevance value takes effect only in semantic search or when reranking is used, and it is a filter threshold that performs no permission or entity isolation.
Hypothetical paper exercise
The following material and request are paper-based, used only to check this page's function. They represent no real record and no completed operation.
Material text (reshipment note, hypothetical): Order BF-2024-0871, product is a portable juicer, model PZ-200, color white, fault is a broken cup latch, support has registered a reshipment of one unit of the same model and color. The order status is currently awaiting warehouse stock confirmation.
Material text (script template excerpt, hypothetical): For inquiries about the PZ-200 latch issue, the standard wording is that a reshipment has been registered and the reshipment timing follows the warehouse dispatch notice.
Draft sentence to be judged (hypothetical): For the refund application concerning the PZ-200 broken latch, reply "reshipment confirmed, no further refund submission needed."
Judgment point: the draft treats a registered reshipment as confirmed and widens the refund handling boundary to no further submission. The material only shows awaiting warehouse stock confirmation and states neither an approved reshipment nor a closed refund application. The draft sentence is unsupported, and the evidence gap must be recorded.
Configuration and implementation checks
The table lists documented settings or implementation checks tied to this page's primary task. Where a number appears, it is a hypothetical starting point for this exercise with a stated comparison target.
| Configuration or check | Hypothetical value or action | Basis and comparison method |
|---|---|---|
| Linked knowledge bases | Select libraries whose vector model identifier matches | The node requires the same vector model; check identifiers first and exclude mismatches |
| Search mode | Hybrid search | Model and order-number queries rely on full-text hits; natural-language questions rely on semantic recall |
| Citation limit | Hypothetical starting point 3000 tokens | The document controls the citation limit in tokens; compare 2000 and 4000 for completeness of cited source text |
| Minimum relevance | Hypothetical starting point 0.4 | Documented as a 0-1 value effective only in semantic search or reranking; compare 0.3 and 0.5 for remaining item counts after filtering |
| Question optimization | Enable for multi-turn follow-up scenarios | Documented as adding one model call and improving continuous-conversation retrieval |
| Rerank | Enable when a ReRank model is available | Documented that rerank scores support filtering; compare ordering with the disabled state |
Three common errors
- Treating a registered reshipment as a confirmed reshipment. Input is the material phrase "awaiting warehouse stock confirmation"; the comparison target is the draft phrase "reshipment confirmed"; the observable result is that the two status fields disagree, and the draft is unsupported.
- Treating retrieval output as an authorization basis. Input is the recalled reshipment note; the comparison target is the refund application handling action; the observable result is that the retrieval chain produces no approval or refund-closure record, and any business action requires a downstream implementation.
- Forcing an association between knowledge bases with different vector models. Input is two libraries with different model identifiers; the comparison target is the node requirement of the same vector model; the observable result is that mismatched identifiers are excluded and no cross-model retrieval comparison is run.
How to verify the setup
- For the hypothetical draft sentence, input the material text and the draft, compare "awaiting warehouse stock confirmation" with "reshipment confirmed," and pass when the draft is marked unsupported with the evidence gap recorded.
- For the hypothetical refund request, input the draft phrase "no further refund submission needed," compare whether the material states the refund application is closed, and pass when the material contains no such field and the result is recorded as pending verification.
- For linked knowledge bases, input two libraries with different model identifiers, compare whether the identifiers match, and pass when the mismatched one is excluded and no cross-model retrieval is executed.
Citation output is an array and its length can be 0. Use an explicit paper input of an empty array to check the downstream design, without claiming that any invented rule always returns empty.