What data looks like for this category
Data sources for aquaculture intelligent due diligence reports include daily culture pond ledgers, online water quality monitoring device data, feed feeding records, seedling purchase vouchers, disease diagnosis and treatment records, and slaughter inspection reports. Data update rhythms vary: Water quality monitoring data generates new records hourly. Daily feeding and ledger information is updated daily. Purchase, diagnosis, treatment, and slaughter vouchers are one-time uploaded static documents. Most documents are structured tables with fields such as culture pond number, monitoring date, dissolved oxygen value, and feeding amount. A small number of handwritten diagnosis records, sick fish photos, and device screenshots are also included. Field units include mg/L (dissolved oxygen), kg (feeding amount/slaughter weight), and tail (seedling quantity). pH values have no unit.
Constraints imposed on model access and configuration
The high share of structured tables and special field units require models to accurately identify and retain field units. Format requirements must be clearly specified in prompts. Mixed image documents require access to multimodal-capable models, with corresponding switches configured. Data update frequencies differ. High-frequency water quality data requires incremental indexing to avoid excessive time spent on full re-runs. One-time uploaded voucher documents have wide size variations. A reasonable upload file limit must be set. Decentralized data from multiple culture ponds requires recall of multiple related entries. The number of recalled entries must be adjusted to match data volume. These characteristics require the model access and configuration process to balance structured parsing, multimodal support, indexing efficiency, and context length control.
Configuration settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | 8000–12000 characters | Aquaculture due diligence data includes multiple sets of structured tables and a small amount of image descriptions. This range aligns with common window limits for mainstream multimodal models and prevents context overflow |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Large culture ledgers may contain multiple pages of associated tables, leading to long parsing times. This duration covers the parsing process for most files |
RERANK_MODEL_ENABLE | Enabled | Structured data recall requires reordering highly relevant fields and entries to improve extraction accuracy for due diligence reports |
RECALL_TOP_N | Top 10 entries | Due diligence reports need to cover core data from multiple culture ponds. This value balances recall completeness and context redundancy |
MULTIMODAL_MODEL_ENABLE | Enabled | Documents include unstructured content such as sick fish photos and water quality device screenshots. Multimodal recognition must be enabled to extract image information |
UPLOAD_FILE_MAX_SIZE | 500 MB | Adapts to large packaged culture ledger files for batch uploads, preventing upload failures due to size restrictions |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require case-by-case analysis. It is recommended to test on your own samples before finalizing settings.
Three common configuration errors
- Phenomenon:
RERANK_MODEL_ENABLEis enabled during online recall testing, but results are not reordered by relevance. Cause: The indexing enable switch for the reranking model was not enabled synchronously in the knowledge base index settings. - Phenomenon: After setting
RECALL_TOP_Nto 3000, the large language model does not receive context. Cause: The total context length exceeds the window limit supported by the model, resulting in automatic truncation or discarding by the system. - Phenomenon: The model cannot recognize image content after multimodal documents are uploaded. Cause: A multimodal-capable model was not accessed, or the
MULTIMODAL_MODEL_ENABLEconfiguration item was not enabled.
How to confirm successful configuration
- Upload a test document containing structured aquaculture tables and sick fish photos, and check if the parsed chunked content includes image descriptions and complete field information.
- Initiate an online recall test, enter a query related to aquaculture data, and verify that recall results are sorted by relevance priority.
- Adjust the value of
RECALL_TOP_Nto confirm that context is properly passed to the large language model's reply stage. - Check system parsing logs to confirm that files were not blocked due to timeout or size restrictions.
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.