What the documents look like
Brand livestream operations place scriptwriting, inventory and after-sales responsibility on the brand itself, and commission structures tied to external hosts do not dominate the work. The documents that matter for contract and transaction material comparison are the livestream script, the product explanation card, the shift schedule and the livestream review, each refreshed per session, while transaction details and livestream comments change by the hour. Comparison work has to place these documents and transaction records inside one retrieval scope so that subject, object, amount and status can be checked against each other.
Fast material turnover creates a direct constraint. A retrieved script version, explanation card version and same-day transaction record may belong to different sessions. Knowledge base search operates at the library level, and collections serve only to classify and manage data, with no effect on search results. Materials held in libraries that use different vector models cannot be associated together.
Constraints at the retrieval step
One retrieval combines question optimization, semantic recall, full-text recall, RRF fusion and reranking, then passes through similarity filtering and reference cap trimming. The output is reference content, and it does not constitute authorization, approval or transaction confirmation. Forms collect values, and business verification, approval, database writes and external synchronization each require their own downstream implementation.
Hypothetical paper exercise sample: the source material reads "Party A, the brand owner, and Party B, the agency operations service provider, signed the Livestream Agency Operations Service Contract on March 1, 2025, agreeing that Party B would handle daily operations of the self-operated livestream room from March to May 2025, with a monthly service fee of RMB 80,000, settled by Party A on a quarterly basis." A second sentence reads "The April 2025 shift schedule shows host A covering evening livestreams from April 10 to April 20, with floor control handled by B." The draft sentence to be judged reads "Party B has completed all agency operations services for April 2025, and Party A shall pay the April service fee of RMB 80,000 by April 30." This draft sentence lacks an acceptance record for April service completion and a settlement trigger condition. The retrieved contract and shift schedule cannot support the conclusions "has completed" and "shall pay," so the draft is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Contract library, shift library, transaction detail library | Check the vector model identifier of each library first, and associate only libraries with matching identifiers |
| Search mode | Hybrid search | Use full-text for contract numbers and amounts, semantic for natural-language phrasing, and compare hit differences between the two modes |
| Reference cap | Hypothetical starting value of 2000 tokens | Test content completeness with contract clauses and transaction records of different lengths, and compare whether amount or subject is lost before and after trimming |
| Minimum relevance | Hypothetical starting value of 0.35 | Label relevant and irrelevant samples first, then compare whether irrelevant fragments are filtered at this value |
| Question optimization | Enabled | Compare the original question with the optimized question, and check whether subject and period were altered incorrectly |
| Result reranking | Enabled | When candidates are numerous, compare ordering before and after reranking, and use the rerank score for filtering |
Three common errors
- Treating a retrieval hit as completed authorization or approval. Retrieving contract text only shows that reference content exists. Signing, reaching cooperation and contract effectiveness are three distinct facts, and a pending review does not demonstrate a lack of operating permission.
- Associating libraries across different vector models. Before association, check only the identifiers and exclude those that do not match. Do not run cross-model retrieval to compare missed recalls.
- Using minimum relevance to filter unauthorized records. Relevance is a filtering threshold and does not carry permission or entity isolation. Retrieval is not authorization, and unauthorized access needs separate control outside the retrieval scope.
How to verify the setup
- Input the source material above together with the draft sentence, and compare whether "has completed" and "shall pay" in the draft are supported by an acceptance record or a settlement trigger condition. When retrieval returns only the contract and shift schedule and does not return an acceptance record, record the result as unsupported, and do not rewrite it as a contradiction or a negation.
- Input a contract number or amount field, and compare the hits from full-text search and hybrid search. If full-text search hits a contract clause while hybrid search does not, record the difference and check the minimum relevance and reference cap settings, without presuming that one mode is necessarily better.
- Input a request that carries only related-stage evidence, such as "April service fee payment application," and compare the request status with the contract settlement clause. When the request is pending review, record it as pending verification, and do not judge it as approved or rejected. An empty result can be checked against downstream design using the paper input []. Do not claim that a search for a fictional rule necessarily returns [].