What the data for this category looks like
Data for aquaculture includes production ledgers, seedling cultivation records, feed feeding logs, disease prevention technical documents, and marketing materials for farmers. These materials include short video scripts, poster copy, and scenario-based talking points. Teams update daily production data each day. They adjust and publish marketing materials as needed. They supplement disease prevention materials with new cases. Documents contain structured aquaculture data and unstructured technical descriptions and marketing content. Fields include pond ID, dissolved oxygen, water temperature, feed model, disease name, as well as applicable aquaculture categories and scenario tags for marketing materials.
What constraints these characteristics impose on knowledge base retrieval and recall
Structured aquaculture data requires precise matching of professional fields. General semantic retrieval cannot cover the precise needs of niche scenarios. Marketing materials vary widely in length. Teams store short talking points and long technical manuals together. Segmentation must retain scenario-related information to avoid losing business context after splitting. Frequently updated production data requires regular synchronization to the knowledge base. Otherwise, recall results will lag behind actual aquaculture conditions. Multi-category, multi-scenario marketing content can lead to generalized retrieval results. Teams must use field filtering to narrow the recall scope. Image-based marketing materials must be bound to associated text to ensure linked display during retrieval.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
segment_length | 800–1200 characters | Aquaculture documents include both short marketing talking points and long technical manuals. This range balances semantic completeness and retrieval accuracy, avoiding loss of scenario association when splitting long documents |
recall_count | Top 8–12 results | Aquaculture marketing content needs to cover multiple dimensions such as breeding scenarios, disease types, and feed selection. Too many results increase screening costs, while too few fail to cover requirements |
similarity_threshold | 0.72–0.80 | There are many professional terms in aquaculture scenarios. This range filters out low-match irrelevant documents while retaining weakly matched results for niche scenarios |
rerank_return_count | Top 3–5 results | Marketing content should prioritize the most relevant talking points or technical solutions for user needs. Limiting the number of results after reranking improves content readability |
PARSE_FILE_TIMEOUT_SECONDS | 300 seconds | Parsing large aquaculture ledgers or batch marketing materials takes a long time. This duration avoids parsing interruptions |
UPLOAD_FILE_MAX_SIZE | 2000 MB | High-definition pond real-world images and long video script documents have large file sizes. This threshold supports batch uploads |
The parameter values provided on this page are conventional recommendations used as a starting point for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require individual analysis, and it is recommended to test on your own samples before finalizing settings.
Three common mistakes
- Phenomenon: After configuring
rerank_return_count, there is no obvious change in retrieval result ranking. Cause: Therecall_countparameter was not adjusted synchronously. The reranking model only applies to the top N recalled results. If the recall count is set much larger than the rerank return count, reranking cannot affect the final display results. - Phenomenon: Pond real-world images and disease comparison images displayed in the knowledge base cannot load normally in retrieval results. Cause: The relative path or external link address of the image was not correctly extracted during document parsing, or the knowledge base configuration did not enable the image-associated retrieval function.
- Phenomenon: When upgrading from version v4.6.7 to v4.8.10, existing knowledge base content was not retained. Cause: Only the container image was pulled without mounting the original data volume, and the vector database and document storage directory of the knowledge base were not backed up.
How to confirm proper configuration
- Upload a test document containing pond data and marketing talking points, check if the parsed segments fall within the expected
segment_lengthrange, and adjust parameters until the segmentation logic meets business requirements. - Enter aquaculture scenario keywords such as "Pacific white shrimp disease emergency response", verify that the number of retrieval results matches the
recall_countsetting, and confirm that the reranking model affects the result order. - Upload a marketing material document containing images, click the corresponding entry in the retrieval results to confirm that the images load and display normally.
- Check the knowledge base upload logs to confirm that batch-uploaded aquaculture ledger documents do not show parsing timeout or failure status codes, verifying that the
PARSE_FILE_TIMEOUT_SECONDSsetting is reasonable.
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.