What this category’s data looks like
Baijiu-related data comes primarily from national distilled spirit standard documents, publicly available process parameter documents from manufacturers, test reports issued by third-party quality inspection institutions, and category development materials released by industry associations. Update rhythms vary across data sources. National standards have long revision cycles. Internal enterprise documents update alongside production process adjustments. Third-party reports update with each inspection batch. Documents mostly use a chapter-based structure, with modules including raw material ratios, brewing processes, physical and chemical indicators, and sensory evaluation. Fields include clear identifiers such as alcohol content (unit %vol), total acid and total ester content (unit g/L), production date, production batch, and inspection item number.
What constraints these characteristics impose on knowledge base retrieval and recall
The multi-source nature of baijiu data requires retrieval systems to distinguish the authority of documents from different sources, and prevent irrelevant content from being included in due diligence reports. The strong binding between fields and their units requires retrieval processes to match the units associated with each field. Failure to do so will produce incorrect results with mixed indicators. Long process documents and detailed physical and chemical indicators require segmented retrieval to retain the connection between indicators and their context, avoiding the separation of critical information. Data sources with different update frequencies require index refreshes to align with each source’s update rhythm, ensuring the timeliness of recalled content.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Baijiu documents include physical and chemical indicators and process steps. Excessively long segments will lose context association, while excessively short segments will separate indicators from their associated scenarios |
chunk_overlap | 150–200 characters | Key baijiu indicators such as alcohol content and total ester content often appear at the beginning and end of paragraphs. Overlapping segments can retain indicator associations across segments |
recall_top_k | Top 8–12 results | Baijiu due diligence requires coverage of multiple sources including standards, enterprise data, and third-party reports. Too many results will increase subsequent processing burden, while too few will miss matching content for specific scenarios |
similarity_threshold | 0.72–0.78 | Baijiu detection indicators require precise numerical matching. A threshold that is too low will introduce irrelevant non-baijiu category data, while a threshold that is too high will miss content with different expressions of the same indicator |
index_refresh_interval | Every 24 hours | Enterprise process documents and third-party test reports are updated on a daily or weekly cycle. Regular index refreshes ensure the timeliness of recalled content |
rerank_top_n | Top 3–5 results | Due diligence reports need to prioritize content from national standards and authoritative test reports. Reranking filters redundant results with low relevance |
The parameter values provided on this page are all conventional recommendations used to establish a starting point for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require specific analysis, and it is recommended to test on your own samples before finalizing values.
Three common mistakes
- Phenomenon: After mounting two knowledge bases, retrieval results only display matching content from a single knowledge base, and cannot return results from both. Cause: The cross-knowledge base recall weight parameter is not configured, and the system loads vector data from only the initially mounted knowledge base by default.
- Phenomenon: After importing knowledge base files via
mongorestore, no matching results are returned during retrieval, and the interface displays "No relevant content found". Cause: The import operation did not trigger the vector index generation process, and the original files were not parsed into retrievable vector data. - Phenomenon: When retrieving baijiu physical and chemical indicators, the returned results contain content with mismatched units, such as confusing "g/L" with "mg/L". Cause: Field-level unit matching verification is not enabled, and retrieval is only based on text semantic similarity without binding field unit parameters.
How to confirm correct configuration
- Perform a single-document retrieval test, enter the baijiu standard number, and check whether the returned results include the core content of the corresponding document.
- Mount two knowledge bases containing baijiu data, enter a query term covering content from both, and confirm that the results come from both knowledge bases.
- View the index refresh log, confirm that the scheduled task configured via
index_refresh_intervalhas been triggered normally with no failed records. - Modify a baijiu test report in the knowledge base, perform a rebuild index operation, then retrieve the new content of the report, and confirm that the results have been updated.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.