What the data for this category looks like
Data sources include public monitoring data from baijiu producing region associations, quarterly financial reports of listed liquor manufacturers, terminal sales data from third-party liquor circulation platforms, and official product specifications and quality inspection reports released by distilleries. Update rhythms: Listed liquor company financial reports are updated quarterly. Producing region association data is updated semi-annually. New product information is updated in real time as released. Circulation data is synchronized daily. A single due diligence document typically includes five modules: region traceability, physical and chemical indicators of liquor body, production capacity breakdown, channel coverage, and competitor benchmarking. Fields include alcohol content (unit: %vol), base liquor storage period (unit: years), single factory production capacity (unit: kiloliters), and revenue scale (unit: ten thousand yuan).
What constraints these characteristics impose on vector models and indexing
Mixed update rhythms require flexible configuration of incremental and full indexing to avoid resource waste from full index rebuilding. Multi-module, multi-field document structures require differentiated chunking for text from different modules, to prevent irrelevant cross-module information from being recalled together. Mixed input of short-text circulation data and long-text financial reports requires adapting chunking strategies for different text lengths, to avoid over-truncating short texts or using unreasonable chunk granularity for long texts. Mixed existence of numeric fields and unstructured text requires supporting fused indexing of structured and unstructured vectors, to ensure recall accuracy.
How to set configurations
| Configuration Item | Recommended Values | Rationale |
|---|---|---|
chunk_size | 800–1200 characters | Adapts to the chunking needs of long-text financial reports in baijiu due diligence reports, while retaining complete semantics of short texts such as circulation data and liquor body parameters |
chunk_overlap | 50–80 characters | Prevents key professional terms across chunks from being truncated, ensuring semantic coherence of content such as baijiu producing regions and liquor body descriptions |
vector_model | text-embedding-3-large or bge-large-zh-v1.5 | Matches the professional terminology of the baijiu industry, including semantic recall accuracy for region classifications and physical and chemical indicators of liquor bodies |
recall_top_k | 10–15 entries | Covers multi-module information required for baijiu due diligence, avoiding missing key data due to too few recalls and introducing redundant content due to too many |
rerank_top_k | 3–5 entries | Streamlines returned results, focusing on core due diligence content such as liquor body parameters, production capacity data, and channel information |
incremental_index_enable | Enabled | Adapts to the mixed update rhythm of baijiu data, reduces resource consumption of full indexing, and meets update requirements for quarterly financial reports and real-time circulation data |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require specific analysis. It is recommended to test on your own samples before finalizing settings.
Three common mistakes
- Symptom: Vector indexing tasks remain in a queued state for a long time with no progress updates. Cause: Incremental indexing is not enabled, and full indexing covers datasets with mixed updates such as quarterly financial reports and real-time circulation data, resulting in excessive resource usage for indexing tasks.
- Symptom: Knowledge base search returns irrelevant historical baijiu producing region text and fails to match core liquor body parameters. Cause: Chunk granularity is set too large, and irrelevant cross-module text is merged into one chunk, leading to deviations in semantic recall.
- Symptom: Search results for numeric fields such as alcohol content and single factory production capacity are empty, and structured data cannot be recalled. Cause: Structured vector fused indexing is not configured, and vectors are only generated for unstructured text, resulting in numeric fields failing to be matched effectively.
How to confirm correct configuration
- View indexing task run logs to confirm that chunk length matches the configured
chunk_size, with no over-truncation or cross-module merging. - Launch search tests for liquor body parameters or production capacity data, check the matching degree between returned result modules and queries, and adjust recall and reranking configuration values.
- Check incremental index trigger records to confirm that updates of quarterly financial reports and real-time circulation data can automatically trigger incremental indexing tasks.
- View vector model call logs to confirm that the structured vector generation function for numeric fields has been properly enabled.
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.