What the documents look like
Livestream after-sales work clusters after a single high-traffic session. Refund requests and reshipment orders arrive in dense batches during the same window, and support staff must give a consistent explanation of what the host promised on air. Routine materials include livestream order lists, refund applications, reshipment orders, and script templates. Sources span shop after-sales records, customer-service IM, warehouse WMS, and logistics tracking. In cross-border work these materials carry Chinese, English, and other languages, so the same fact appears under different wording across languages and retrieval must cope with that variation.
The update rhythm given here is per session and daily. That rhythm affects when material becomes retrievable; the retrieval step itself runs per request and does not automatically trigger full or real-time synchronization because of the update frequency.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together. Check the vector model identifier first, exclude those that do not match, and record what was observed; do not run cross-model retrieval to compare missed recalls.
Search mode is compared per test question across semantic retrieval, full-text retrieval, and hybrid retrieval. No single mode is presumed suitable for all material. Semantic retrieval suits natural-language phrasing; full-text retrieval suits numbers, models, and proper nouns; hybrid retrieval merges both paths through RRF. RRF is the merge process for hybrid results and is not described as a standalone switch. The citation limit is measured in tokens and controls how much content a single search may cite at most. Minimum relevance takes effect only under semantic retrieval or when result reranking is used, and it is a filter threshold. Question optimization performs coreference resolution and question expansion and adds one model call. Image retrieval requires corresponding image input and model capability; reranking requires a model and must be enabled by setting. The optional paths listed in the overall flow diagram do not mean every request executes all of them.
Retrieval output is an array of cited content whose length can be 0; when nothing is found, the output link still runs through. Retrieval and recall do not perform authorization judgment, and a similarity score does not constitute permission or entity isolation.
Hypothetical exercise sample: the source text reads, "Order A-2201 submitted a refund application on March 5, status pending review; the host stated during the March 4 livestream that this batch could be reshipped." The draft sentence to be judged reads, "The refund for order A-2201 has been approved and the reshipment has been arranged." The source only records that the refund application is pending review, not that it was approved; the reshipment statement does not record that execution was arranged.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate one knowledge base using the same vector model | Associate after checking that vector model identifiers match; exclude cross-model bases and do not run cross-model retrieval comparison |
| Search mode | Test hybrid retrieval first | Compare semantic, full-text, and hybrid results separately using an "A-2201"-type number question and a natural-language question from the sample |
| Citation limit | Hypothetical starting point of 2000 tokens | Test whether content is truncated using refund applications and script templates of different lengths, and compare completeness against a smaller value |
| Minimum relevance | Hypothetical starting point of 0.5 | Label relevant and irrelevant samples first, then compare whether irrelevant fragments decrease under different thresholds |
| Question optimization | Off first, enabled for multi-turn follow-up scenarios | Compare the original question with the optimized question, check whether the subject was altered, and record the change in elapsed time |
| Result reranking | Enabled when candidates are numerous | Compare the ranking position of the same question before and after reranking; the rerank score is used for filtering |
Three common errors
- Treating a retrieval hit as authorization approval. In the sample, the refund application status is pending review. Retrieving that application only shows the material exists; it does not show the refund was approved or the reshipment executed. To check, input that application number, compare the status field in the retrieval result with the wording of the draft sentence, and treat "approved" or "arranged" in the draft as inconsistent with the source.
- Treating a similarity score as an entity or permission filter. Minimum relevance only filters low-relevance results; it does not exclude unrelated business or unauthorized records. To check, input a question unrelated to after-sales, compare whether returned content can still include other order material, and note that inferring permission isolation from this lacks a basis.
- Associating knowledge bases with different vector models directly. The documentation requires associating knowledge bases that use the same vector model. To check, verify each base's vector model identifier first, exclude those that do not match, record this, and do not continue comparing retrieval results.
How to verify the setup
- Input the sample source text and the draft sentence, and compare the refund status and reshipment status in the draft with the source's recorded "pending review" and "stated could be reshipped." The draft fails if it says approved or arranged, and should be changed to pending review and promised but not yet executed.
- Input the order number "A-2201" and the natural-language question "what is the status of this refund now," and compare whether the cited content returned under both phrasings includes that order's source text. If the number phrasing misses while the natural-language phrasing hits, record that difference and check the search mode setting; do not conclude from this what caused a missed recall.
- Use an explicit paper input of an empty array to check the downstream design, confirming that when cited content length is 0 the output link still runs through and downstream can accept the empty result; do not claim that a fabricated rule's retrieval necessarily returns an empty array.