What the documents look like
Livestream after-sales work arrives in concentrated bursts after a single broadcast session. Refund requests and replacement orders enter at the same time, often within a short window. The material sources named in the brief are the shop after-sales system, customer service IM, warehouse WMS, and logistics tracking. The document types are livestream order lists, refund requests, replacement orders, and script templates.
These materials need to be grouped by broadcast session. For one order, the commitment made by the host or customer service agent, the refund request, and the replacement arrangement sit in different documents. The chunking task is to split these materials into retrievable units so that contract and transaction material comparison can locate a specific order and a specific commitment.
The brief marks its confidence as broadly accurate. The timing pattern and the document list are the parts that shape this page. Update frequency is per session and daily, and it only matters here because chunking must preserve fields that stay stable within a session.
Constraints at the document chunking step
Chunking must keep order number, session, commitment subject, and status field together. If a condition and the content it constrains are split apart, a later comparison step can match a fragment that no longer carries its own scope.
FastGPT stores original data and vector indexes separately. The dataset.datas table in MongoDB holds the original data and an indexes field that records vector IDs as an array, so one group of original data can map to multiple vectors. The PostgresSQL table holds a vector field. During retrieval, vectors are recalled first, then the vector ID is used to find the original content in MongoDB; results that map to the same original data are merged and the highest vector score is kept.
A successful upload does not mean training and index update are complete. The chunking step therefore has to treat upload status and training status as two separate facts.
Hypothetical drill sample. The source material reads: "Order number HX-2024-0731-A12, purchased a skincare set in the livestream session on 31 July 2024; customer service promised a replacement of the missing sample within 7 days. Refund request R-0731-056 shows this order applied for refund only, reason: missing item." The draft sentence to be judged is: "This order has arranged the replacement, and the refund request should be closed."
This sample is constructed for this page. It is not a real order, not a real customer, and not a real authorization record.
Configuration and implementation checks
| Configuration item | Setting or check relevant to this page | Relation to the sample |
|---|---|---|
| Search mode | Semantic retrieval, full-text retrieval, or hybrid retrieval | Order number uses full-text retrieval; commitment description uses semantic retrieval |
| Reference limit | Reference limit measured in tokens | Controls the length of reference content returned to the model |
| Minimum relevance | A 0-1 value, effective only under semantic retrieval or result reranking | Filters low-relevance fragments; does not act as a permission filter |
| Question optimization | Coreference resolution and question expansion before retrieval | Completes order references in multi-turn follow-up |
| Result reranking | ReRank model performs a second ordering, score 0-1 | Improves text relevance when candidates are numerous |
| Data and index | Check original data and index status | Confirm training completion before retrieval |
Hypothetical trial starting value for this page: minimum relevance set to 0.5. The reason for choosing it is that the sample question is explicit and the candidate fragments are limited. The comparison objects are the recalled content under 0.3 and under 0.7. This value is a hypothetical starting point for this drill and is not an official default.
Three common errors
- Treating a successful upload as an updated index. Input: a replacement order that was just uploaded. Compare the upload status with the actual training status. If retrieval cannot find that replacement order content, the index has not taken effect.
- Using a similarity score to exclude unrelated business records. Input: order number HX-2024-0731-A12. Compare whether the returned fragments contain that order number. If the fragments return similar commitment descriptions from other orders, semantic recall has not substituted for entity filtering.
- Treating a draft sentence awaiting judgment as a confirmed fact. Input: the draft sentence "this order has arranged the replacement." Compare whether the source material contains a replacement execution record. The source material states only a customer service promise and does not state that the replacement was executed, so the sentence is unsupported.
How to verify the setup
- Input the sample source material and the draft sentence. Compare "has arranged the replacement" in the draft against "customer service promised a replacement within 7 days" in the source. The source does not state a replacement execution action, so the judgment is unsupported and the actual status of the replacement order remains to be verified.
- Input order number HX-2024-0731-A12 for full-text retrieval. Compare whether the returned fragments contain that order number and the corresponding refund request R-0731-056. If the returned fragments do not contain that order number, record it as a miss and check the index status.
- Input an empty-result paper input []. Compare how the downstream design handles empty references, and observe whether it clearly distinguishes no result from not retrieved. Do not presume that a fictional rule returns an empty result.
Retrieval checks record actual observations and failure branches. They do not guarantee a first-position hit or full recall. Libraries built on different vector models cannot be associated together; check the model identifier first, exclude those that do not match, and record the actual observation.
This page covers only livestream after-sales concentrated processing, contract and transaction material comparison, and document parsing and chunking. During operational handover, the authorization status of materials, operations, and requests cannot be treated as authorization execution through retrieval or a form.