What the documents look like
Livestream after-sales work concentrates after a single broadcast. Refund requests and replacement orders arrive in large volumes within a short window, and the handling basis is closely tied to what the host promised on air. Inventory procurement and supplier records therefore support two kinds of lookup: the procurement source, supplier contact and arrival record for the item named in a refund or replacement, and the original wording of a commitment recorded in a script template. Data sources include the shop after-sales system, customer service IM, warehouse WMS and logistics tracking. Everyday documents are livestream order lists, refund requests, replacement orders and script templates.
The retrieval step is shaped by the form of the material. Supplier names, procurement order numbers and model numbers are exact fields, suited to full-text retrieval. Refund reason descriptions and explanations of commitments are natural language, suited to semantic retrieval. One item can correspond to several supplier records, and a single broadcast's order list and replacement orders need to be traced by order number. Retrieval term quality directly affects whether the correct material can be located.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. The smallest search unit is the knowledge base; a collection is only a way to classify data and does not affect search results. Semantic retrieval relies on vector distance, full-text retrieval relies on keyword hits, and hybrid retrieval merges the two result sets through RRF. Re-ranking requires a re-rank model and must be enabled according to settings. Image retrieval requires corresponding image input and model capability. These optional paths do not mean every request executes all of them.
The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result re-ranking is used. Both are filtering and trimming conditions. The relevance score returned by retrieval is used for ordering and filtering; it does not perform permission or entity isolation, and retrieval itself does not constitute authorization. Question optimization performs reference resolution and question expansion, and adds one model call before formal retrieval.
Hypothetical drill sample: after a livestream session on 18 June 2025, a refund request is received. The request form names the item as "Type A folding table", procurement order number "CG-20250610-07", supplier "Example Furniture Manufacturing Plant", and the arrival record shows 20 units received into the warehouse on 12 June 2025. The original script template states: "This batch supports seven-day no-reason return." The draft sentence to be assessed is: "This batch does not support return; please guide the customer to choose a replacement." The object of comparison is the original script template wording and the draft sentence, checking whether the two statements about the return condition agree.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Inventory procurement base, supplier records base | Both bases use the same vector model; associate after checking the model identifier |
| Search mode | Hybrid retrieval | Procurement order numbers need full-text hits, refund reasons need semantic recall; compare the two result paths |
| Citation limit | Hypothetical starting point 3000 tokens | Used to hold procurement order and arrival record fragments; compare against the actual completeness of shorter material |
| Minimum relevance | Hypothetical starting point 0.3 | First label relevant and irrelevant samples, then compare what is retained after filtering |
| Question optimization | Enabled | Completes the subject in follow-up turns; compare whether the subject in the original question and the optimized question agree |
| Result re-ranking | Enabled according to settings | Second ordering when candidates are numerous; compare the re-rank score with the vector score |
These values are hypothetical starting points for this exercise and do not come from official default values. The configuration covers only the main task on this page; internal database implementation is not listed item by item as a required setting.
Three common errors
- Treating the relevance score as a permission filter. Retrieval returning a supplier record does not mean the current customer service account is entitled to view that supplier's procurement prices or contract terms. To check, enter the same query, compare the returned content with the range accessible to the current account, and observe whether material beyond that range appears.
- Forcing an association across knowledge bases with different vector models. When two bases have different vector model identifiers, the node cannot associate them together. To check, verify the model identifiers, exclude those that do not match, and record the actual observation. Do not run a cross-model retrieval to compare missed recalls.
- Treating a pending request as already authorized. When a refund request is in pending review status, the draft sentence cannot state that the return has been approved. To check, enter the request status field, compare the status value with the wording of the draft sentence, and observe whether status and wording disagree.
How to verify the setup
- For the script template original text and the draft sentence in the hypothetical sample, enter the draft sentence for retrieval, compare whether the returned template original contains "supports seven-day no-reason return", and observe whether the draft sentence contradicts the original. The expectation is that the template fragment can be located; if the returned content does not contain that sentence, record it as a miss.
- For procurement order number "CG-20250610-07", enter the number for full-text retrieval, compare the procurement order number field in the returned result, and observe whether it matches exactly. The expectation is a hit on the corresponding procurement record; if a different order number is returned, record it as a field mismatch.
- For the empty-citation case, use the explicit paper input [] to check the downstream design, compare whether the node output length is 0, and observe whether the subsequent chain still runs through. The expectation is an empty array output with the chain continuing. Do not claim that a fabricated rule retrieval necessarily returns an empty result.
All three checks use the hypothetical sample as input and do not indicate that any check has actually been executed. Fields not stated remain unknown, and a pending-verification status is not judged as a negative.