What the documents look like
Group-buy leader operations centre on community pickup nodes and commission settlement. The working materials are leader agreements, commission tables, pickup lists and community announcements. Each material carries a different subject, object, validity range and status. An agreement defines the scope of cooperation. A commission table constrains how settlement is calculated. A pickup list corresponds to a specific group round and quantity. A community announcement addresses neighbours. Product information and selection questions need to retrieve citable fragments from these materials, so chunking has to preserve material identity and clause boundaries. Materials update daily or per group round. After a change, the actual training and index state needs checking. A successful upload does not mean retrieval has updated.
The knowledge base is made of libraries, collections and data. The smallest search unit is the library; collections only classify and manage data and do not affect search results. Vectors are stored in a PostgresSQL vector field with an HNSW index. MongoDB's dataset.datas table stores the original data, and its indexes field records the corresponding vector IDs, so one group of data can map to multiple vectors. Retrieval first recalls vectors, then looks up the original data in MongoDB by vector ID. If the same group of original data is matched, the results merge and the highest score is taken.
Constraints at the document chunking step
Segmentation should keep semantics complete and single. Conditions, explanations and the content they constrain must not be split apart. Semantic search depends on vector distance, full-text search depends on keyword hits, hybrid search merges through RRF, and ReRank performs a second ordering that yields a 0-1 score. The minimum relevance value takes effect only under semantic search or result reranking. It is a filter threshold and does not perform permission or entity isolation.
A relevance score is not a category, entity or permission filter. It does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization. Forms collect values; business verification, approval, database writes and external synchronization require their stated downstream implementation.
Document parsing and chunking for product questions must keep the material identity and clause boundaries of leader agreements, commission tables, pickup lists and community announcements. The following hypothetical paper case is used only for this page's verification.
Hypothetical paper test case. Source text: A platform and a group-buy leader agree that from 1 March 2025 to 31 August 2025 the leader is responsible for a pickup point in a named residential community, and commission is settled at five percent of the pickup list amount for the month. Draft sentence to be assessed: The leader still settles commission at five percent in September 2025. The assessment requires the same subject, the same period and the same field. The source period ends on 31 August. September settlement is not supported by the source text and remains pending verification. It does not constitute a negation or a conflict.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting value | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid search | Commission ratios and dates need both semantic and keyword paths; compare candidates after RRF merging |
| Reference limit | Set to 2000 tokens for this trial | Controlled by tokens; compare whether references cover the clause period at different values |
| Minimum relevance | Set to 0.4 for this trial | Takes effect only under semantic search or reranking; compare candidate counts before and after filtering |
| Question optimization | Enabled for this trial | Multi-turn follow-ups need coreference resolution; compare the search terms before and after completion |
| Result reranking | Enabled for this trial | Second ordering when candidates are numerous; compare rerank scores with vector scores |
| Index check | Manual review of the indexes field | Confirm original data maps to vector IDs; exclude cross-model libraries |
The reference limit uses the documented unit of tokens. It is not a guessed item count. Same-model retrieval association, empty-reference output and context limits retain their documented conditions. An unprovided document field is unknown.
Three common errors
- Treating a source agreement whose period ends on 31 August as proof that September settlement is negated. The check takes the draft sentence and the source period field as input, compares whether the two periods match, and observes that the result is pending verification.
- Treating minimum relevance as a permission or entity filter, assuming a low score excludes unrelated leaders. The check takes the similarity setting description as input, compares its conditions of effect, and observes that it only filters under semantic search or reranking.
- Treating a successful upload as retrieval having updated. The check takes the upload record and the training and index state as input, compares whether the two are synchronized, and observes that retrieval still returns old fragments when the index has not updated.
How to verify the setup
- Take the source text and the draft sentence above as input, compare the subject, period and commission ratio fields for the same origin, and judge a contradiction only when all three fields are consistent and their values are mutually exclusive. Otherwise record the result as pending verification.
- Search with a term, an identifier and a natural-language question separately, compare whether the returned fragments preserve the clause period and subject, and observe whether segmentation has split conditions away from the content they constrain.
- Take an empty reference result [] and check the downstream design, comparing whether the downstream handles it as an empty result. Do not claim that a fictional rule search necessarily returns []. For libraries across different vector models, check the identifier first and exclude those that do not match. Do not run cross-model retrieval to compare missed recalls. Image retrieval requires an image input and model capability. Reranking requires a model and must be enabled according to settings.