What This Category of Data Looks Like
Aquaculture investment research data comes from multiple sources. These include on-site pond monitoring data collected at aquaculture terminals, publicly available aquaculture technical documents from industry associations, satellite remote sensing water environment images, and internal log records from aquaculture enterprises.
Data update schedules vary. On-site monitoring data updates daily. Industry technical documents update quarterly. Remote sensing images update weekly.
Individual documents are mostly structured parameter sets or unstructured technical description text. They include fields such as pond ID, start date of the breeding cycle, water temperature, dissolved oxygen, pH value, and feed dosage. Water temperature is measured in degrees Celsius. Dissolved oxygen is measured in milligrams per liter. Feed dosage is measured in kilograms.
What Constraints Do These Characteristics Impose on Retrieval and Recall
Multiple data sources with varying update schedules require the retrieval system to support configuring different update cycles per data type. This prevents stale data from being included in real-time query results.
Documents contain both structured monitoring fields and unstructured technical text. This requires setting differentiated segmentation rules for different document types. This avoids breaking semantic associations of structured fields when splitting long text.
Aquaculture data has strict unit standards. A unified unit conversion logic must be used during retrieval. This prevents recall result deviations caused by unit mismatches.
Highly real-time pond monitoring data requires the recall pipeline to support low-latency incremental retrieval. This ensures the timeliness of query results.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
ENABLE_DOCX_IMAGE_OCR | Enabled | Aquaculture documents often include photos of on-site pond water quality monitoring and images of disease symptoms. Enabling this setting extracts text from images for retrieval purposes. |
maxContext | 800–1200 characters | Individual aquaculture monitoring documents are mostly structured parameter sets. Excessively long context introduces irrelevant fields. This range retains core parameters and associated technical descriptions. |
| Recall Count | Top 6–8 results | Most aquaculture retrieval needs target precise matching of pond parameters or specific disease solutions. A small number of precise recalls avoids redundant results. |
| Similarity Threshold | 0.72–0.80 | Aquaculture data fields have strong correlations. A lower threshold will introduce mismatched pond data. This range balances recall precision and coverage. |
| Maximum Knowledge Base References | Top 3–4 results | Single queries involve limited aquaculture scenario parameters. A small number of references ensures responses focus on the current pond or disease issue. |
PARSE_FILE_TIMEOUT_SECONDS | 120 seconds | Aquaculture documents often include multiple high-definition monitoring images. OCR processing takes longer. This duration prevents parsing timeouts for large documents. |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material format, data volume, and business rules. Each scenario requires specific analysis. It is recommended to test on your own samples before finalizing settings.
Three Common Misconfigurations
- Observation: AI-generated responses consistently include system introduction text that does not appear in the knowledge base. Cause: No exclusive system prompt is configured, or the platform's default built-in introduction text is not overwritten.
- Observation: After deploying via Docker Compose, restarting the container causes loss of knowledge base and workflow configurations, but the API interface responds to requests normally. Cause: The database storage directory is not mapped to a host persistent path. Temporary storage data is cleared when the container restarts.
- Observation: After uploading a DOCX document with photos of on-site pond monitoring, retrieval results do not include text from the images. Cause: The
ENABLE_DOCX_IMAGE_OCRconfiguration item is not enabled, and OCR text extraction for images inside the DOCX file is not performed.
How to Verify Successful Configuration
- Upload a DOCX document containing photos of on-site pond water quality monitoring. Wait for the parsing task to complete. Search for the keyword "pond dissolved oxygen". Confirm retrieval results include content from the uploaded document.
- Configure an exclusive system prompt. Run a test query. Check that the generated response does not include unplanned fixed text at its end.
- Adjust the
Similarity Thresholdparameter. Run multiple test queries with varying precision requirements. Confirm that the matching precision of recall results meets current business needs. - Review file modification logs in the database persistent directory. Confirm that knowledge base incremental update tasks execute at the preset schedule.
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.