What the documents look like
Livestream after-sales work arrives in bursts after a single high-traffic session, and refund and reshipment requests are closely tied to what the host promised on air. For the inventory procurement and supplier materials task, four document types typically enter the knowledge base: livestream order lists, refund applications, reshipment orders, and script templates. The supplied brief lists update frequency as after each session and daily. These documents carry order numbers, supplier names, product specifications, commitment clauses, and validity ranges. Parsing and chunking must preserve those elements so that later retrieval can point back to the original text.
The brief lists the data sources as Douyin store after-sales, customer service IM, warehouse WMS, and logistics tracking. The parsing stage handles only document content that has already entered the knowledge base. It does not replace the business judgment performed inside those systems. The brief marks its confidence level as roughly accurate, so the category description stays conditional rather than fixed.
Constraints at the document chunking step
FastGPT stores knowledge in three parts: library, collection, and data. The smallest search unit is the library, and a search runs across the whole library. A collection classifies data for management and does not affect search results. In the MongoDB dataset.datas table, each data record stores original information and an indexes field that records the vector IDs it maps to. This is an array, so one data record can correspond to multiple vectors. Multiple vector mapping lets one data group map to several vectors, and retrieval recalls the group when any one of those vectors is hit. Chunking therefore needs to follow the actual structure of supplier materials, keeping procurement conditions, reshipment scope, and the constrained products together.
Full-text retrieval suits numbers, models, and proper nouns. Semantic retrieval suits natural-language descriptions. Hybrid retrieval uses both and merges the two result sets through RRF. The minimum relevance score is a filter threshold that takes effect only under semantic retrieval or result re-ranking. It does not perform supplier or permission filtering. The citation limit controls how many tokens of content each search cites. The document describes optional paths such as image retrieval and re-ranking, but those paths require the corresponding model capability and configuration, and they do not run on every request. Libraries built on different vector models cannot be associated together; check the model identifier first and exclude non-matching ones, then record what is observed.
A hypothetical paper exercise sample: the source text reads "Supplier A, March 2025 reshipment order, product number SP-0417, reshipment quantity 20 units, applicable order scope limited to livestream orders from March 1 to March 15. This reshipment order is in pending warehouse confirmation status." A draft sentence to be checked reads "Supplier A has agreed to bear reshipment for all March livestream orders." Checking compares the draft sentence's order scope against the source limit of March 1 to March 15, and compares the source status of pending warehouse confirmation against the draft claim of agreement. The source text does not state that agreement.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | Materials contain numbers and natural-language commitments; compare semantic and full-text recall coverage |
| Minimum relevance score | 0.3 | Hypothetical starting point; compare against 0.5 to observe candidate count after filtering |
| Citation limit | 2000 tokens | Hypothetical starting point; compare against 3000 for citation completeness |
| Question optimization | Enabled | After-sales follow-ups often omit the subject; compare query completeness before and after |
| Result re-ranking | Enabled | When candidates are numerous, compare ranking positions before and after re-ranking |
| Chunking basis | Segment by reshipment order and commitment clause | Compare against whole-document segmentation to see whether conditions and constraints stay in one chunk |
These values are hypothetical starting points for this exercise, not official defaults. Image retrieval requires image input and the corresponding model capability. This page's sample contains no images, so that path stays disabled.
Three common errors
- Treating the minimum relevance score as a supplier or permission filter. Input a record with low relevance that belongs to a different supplier, compare whether it still enters the candidate set, and observe that the relevance setting does not exclude that record.
- Treating a successful upload as retrieval already updated. Input the updated reshipment order number, compare the actual training and index status, and observe whether retrieval results still point to the old version.
- Drafting a reshipment order in pending warehouse confirmation status as already agreed. Input the draft sentence above, compare the source status field, and observe that the claimed agreement is not stated in the source text.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence. Compare the order scope and status field. The check condition is met when the draft sentence matches the source scope and does not rewrite the status.
- Input reshipment order number SP-0417. Compare the results returned by full-text retrieval and semantic retrieval. The condition is met when the number is accurately recalled. When it is not recalled, record the outcome as pending verification.
- Input an empty-result continuation scenario. Use a paper input of [] to check the downstream design and observe whether the downstream handles it as an empty citation. Do not claim that a fictional rule search necessarily returns [].
Libraries across different vector models cannot be associated together. During checking, first inspect the model identifier, exclude non-matching ones, and record the actual observation. Retrieval does not perform authorization judgment. A form collecting values does not mean business verification or a database write has completed. A pending request is not approval, an offered option is not execution, and an unsigned termination is not a completed termination.