What the documents look like
Platform store ratings and experience score maintenance deals with consumer feedback after post-sale fulfilment. The main data comes from Douyin Store experience scores, Tmall service indicators, JD post-sale reviews and customer service systems. Typical documents include negative review handling records, experience score breakdowns, service diagnostic reports and rectification ledgers. A shared characteristic of these materials is that within one store and one indicator period, multiple records reference each other. A single negative review handling record often points simultaneously to an order, a customer service conversation and a rectification entry. Materials update daily or weekly. Parsing and chunking must ensure that indicator names, statistical periods and store entities can still be traced back to the original document after splitting. Otherwise, retrieved fragments cannot support conditional judgement in selection questions.
Constraints at the document chunking step
In FastGPT, a knowledge base consists of library, collection and data. The smallest search unit is the library. Collections only classify and manage data, and have no relationship with search results. Vector source data is stored in the MongoDB dataset.datas table. The indexes field records corresponding vector IDs, and one group of data can correspond to multiple vectors. Vectors are stored in the PostgresSQL vector field. During retrieval, vectors are recalled first, then the original data is fetched from MongoDB by vector ID. If the same group of original data is matched, it is merged and the highest score is taken. When segmenting, conditions, explanations and the content they constrain should remain in the same data block. Splitting the experience score definition from its exception conditions across different blocks should be avoided.
Hypothetical paper exercise sample: the source material text reads "Sample Store A March 2025 experience score breakdown: 1 March to 31 March, the logistics experience score definition covers dispatch timeliness rate and pickup duration. Negative review handling record number C-0312 belongs to this period, status pending customer service supervisor review." The draft sentence to be judged reads "Sample Store A March 2025 negative review handling record C-0312 has completed rectification and been closed." The subject, object and period of the draft sentence match the material, but the material only states pending review. It does not state that rectification was completed or that the record was closed. This is unsupported.
Configuration and implementation checks
| Configuration or check item | Value or status | Relationship to this page's materials |
|---|---|---|
| Search mode | Hybrid retrieval | Experience score definitions contain proper nouns; negative review numbers require exact matching |
| Reference limit | Hypothetical trial starting point 3000 tokens | Negative review records and diagnostic reports differ greatly in block length; controlling by tokens is more stable than top k; compare with 2000 and 4000 starting points for reference completeness |
| Minimum relevance | Hypothetical trial starting point 0.4 | Takes effect only in semantic retrieval or with result reranking; compare with 0.3 and 0.5 for false recall and missed recall |
| Question optimisation | Decide whether to enable based on actual effect | Continuous follow-up asking about the status of this number requires coreference completion; enabling adds one model call |
| Segmentation check | Manual review | Confirm indicator period, store entity and number are not split apart |
| Training status check | Manual review | After material changes, check actual training and index status; successful upload does not mean retrieval has updated |
Three common errors
- Treating successful upload as retrieval already updated. The write status of
dataset.datasandvectormust be checked separately. Form submission alone cannot determine whether an index has been generated. - Writing pending review as completed rectification. The sample material only states "pending customer service supervisor review". The draft sentence writes "has completed rectification and been closed". This treats unsupported content as confirmed.
- Using a similarity threshold in place of entity and period filtering. Minimum relevance is a filtering threshold. It does not perform store or period isolation. Records from another store may be recalled because of semantic similarity.
Corresponding checks:
- Input the sample source material text and the draft sentence. Compare whether "pending review" and "completed rectification and been closed" are mutually exclusive values in the same field. If the latter appears, judge it unsupported.
- Input negative review number C-0312. Use full-text retrieval and semantic retrieval separately. Observe whether the number hits the same data block. If it does not hit, check whether segmentation has split the number away.
- Input experience score breakdowns from the same store in different periods. Observe whether returned blocks carry a period identifier. If the period identifier is missing, record a segmentation gap.
How to verify the setup
Use the sample source text and draft sentence to perform three verification steps.
- First, input "Sample Store A March 2025 negative review handling record C-0312 status". Compare whether the returned block contains "pending customer service supervisor review". If it contains that status, the expectation is met. If it returns "completed rectification", judge it a failure.
- Second, input "experience score logistics experience score definition". Compare whether the returned block is in the same block as the March period. If it is in the same block, the expectation is met. If another period is mixed in, re-segment.
- Third, input an empty query
[]to check downstream design. Observe whether downstream handles it as an empty result. Do not assume a fictional rule must return empty. Reranking requires a model and must be enabled according to settings. Image retrieval requires image input and corresponding model capability. Neither executes by default because of this page's configuration. Libraries across different vector models cannot be associated together. Check the library identifier first and exclude those that do not match. Record actual observations. Do not infer missed recall or redundant results.
This page addresses only: platform store ratings and experience score maintenance / product materials and selection questions / document parsing and chunking. During operational handover, which materials, operations and requests are in which authorisation state cannot be treated as authorisation execution through retrieval or forms.