What the Data for This Category Looks Like
Data sources for crop farming intelligent due diligence reports include farmland soil test reports, crop growth monitoring records, satellite remote sensing imagery, planting ledger Excel files, plant protection service documents, purchase contracts, and more.
Data update schedules adjust with planting cycles. Multiple phased records are generated for each crop from sowing to harvest. Cross-season data is archived annually.
Document structures mix structured tables and unstructured text, including fields such as geographic coordinates, crop varieties, sowing dates, irrigation volume, pesticide application volume, and others. Some fields have dedicated units, such as cubic meters per mu and kilograms per hectare.
What Constraints These Characteristics Impose on Vector Models and Indexing
The multi-source, mixed nature of crop farming due diligence data requires vector models to support vectorization of both text and image materials. This avoids missing key due diligence materials like satellite images and growth photos that text-only models overlook.
Data updates follow planting cycles, creating high demand for incremental updates. The indexing system must support incremental insertion to reduce storage and computing costs.
Fields carry dedicated units, so vector indexes must retain unit information to prevent confusion between indicators from different plots. The system must also support precise matching of structured fields.
A large share of long-text monitoring reports means chunking must preserve context association between adjacent plots or cycles. This avoids semantic fragmentation in the processed data.
How to Configure Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
EMBEDDING_MODEL | clip-ViT-L-14 or bge-large-zh-v1.5 | Supports multi-modal data vectorization, compatible with satellite remote sensing images, crop growth photos and text-based due diligence materials |
SPLIT_CHUNK_SIZE | 800–1200 characters | Crop farming due diligence reports often contain long monitoring records and plot ledgers; this range can retain complete business context for a single planting area |
RECALL_TOP_K | Top 8–12 results | Due diligence requires covering multi-dimensional data including soil, irrigation, plant protection and other aspects; too few recall results will miss key indicators, while too many will increase retrieval latency |
RERANK_MODEL | bge-reranker-large | Improves the accuracy of recall ranking for structured fields and indicators with units, and filters low-relevance retrieval results |
UPLOAD_FILE_MAX_SIZE | 2000 MB | Supports batch upload of large-volume due diligence documents such as satellite image packages and quarterly monitoring collections |
PARSE_IMAGE_ENABLE | Enabled | Compatible with non-text due diligence materials such as crop growth photos and remote sensing images, and generates corresponding vector indexes |
The parameter values provided on this page are all conventional recommendations used as a starting point for configuration. Actual values are affected by material form, data volume and business rules. Specific issues require case-by-case analysis, and it is recommended to test on your own samples before finalizing the configuration.
Three Common Configuration Mistakes
- Symptom: When deploying version 4.9.0 of the knowledge base locally, there is no image indexing option after uploading files, and the vector generation status shows failure. Cause: No multi-modal embedding model is configured, or the
PARSE_IMAGE_ENABLEswitch is not enabled in the knowledge base settings. - Symptom: The embedding model call returns a
500 Internal Server Error, and the logs show connection timeout. Cause: The embedding model address is not correctly mapped in the transit interface configuration, or the corresponding network port is not open for local deployment. - Symptom: After enabling the rerank model, the online recall test results show no sorting changes. Cause: The reranking function is not enabled in the knowledge base index settings, or the rerank model is incompatible with the currently used embedding model.
How to Verify Proper Configuration
- Enter the model configuration page of the knowledge base, check the available embedding model list, and confirm that the target configured model type is included.
- Upload a single crop growth photo, wait for parsing to complete, then check the vector generation logs to confirm there are no error messages.
- Perform a recall test, enable and disable the rerank model separately, compare the result rankings of the two recalls, and confirm that the ranking has changed.
- Check the local deployment environment variables, confirm that the
UPLOAD_FILE_MAX_SIZEconfiguration value is greater than the volume of commonly uploaded documents.
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.