Knowledge Base Retrieval and Recall for Aquaculture Intelligent Due Diligence Reports

Data sources for aquaculture intelligent due diligence reports include real-time collection data from on-site aquaculture monitoring equipment, daily

What this category’s data looks like

Data sources for aquaculture intelligent due diligence reports include real-time collection data from on-site aquaculture monitoring equipment, daily breeding records, seedling breeding logs, feed supply documents, fishery administration spot check reports, and quarterly disease early warning information. Data update frequencies vary: real-time water quality and feeding data updates every 5 minutes. Daily breeding records are updated at midnight each day. Quarterly market and disease early warning data syncs once per quarter. Individual due diligence report documents are organized around breeding batches as the core unit, and include fields such as breeding pond number, monitoring time, dissolved oxygen concentration, water pH value, aquaculture water temperature, daily total feeding amount, seedling release batch, proportion of surviving individuals to total released amount, and total breeding cycle duration. Field units include milligrams per liter, degrees Celsius, kilograms, and days. Each document typically includes summary data for 1 to 3 consecutive breeding cycles and corresponding original records.

What constraints these characteristics impose on knowledge base retrieval and recall

First, high-frequency updates of real-time monitoring data require the knowledge base to support incremental indexing. This avoids performance loss from full re-parsing, which would otherwise fail to ensure the timeliness of retrieved data. Second, documents are organized around breeding batches and include multiple fields with specific units. If retrieval does not split segments by batch or field dimensions, monitoring data from different breeding cycles will be mixed, reducing recall accuracy. Third, fields in due diligence reports have strong correlations. For example, dissolved oxygen concentration must be tied to the corresponding breeding pond and monitoring time. If retrieval does not retain field association context, data matching errors will occur. Fourth, mixed multi-source data requires the retrieval system to support precise cross-data-source matching. This prevents recalling non-corresponding aquaculture data from different sources. Additionally, individual documents have large content volumes. Unreasonable segment rules will cause semantic fragment breaks, affecting large model understanding of complete breeding cycle data.

Configuration Settings

Configuration ItemRecommended ValueRationale
maxContext8000–12000 charactersSingle aquaculture due diligence report documents often include multi-batch breeding data. Longer contexts preserve complete field associations for single batches and avoid interference from cross-batch information
recall countTop 6–8 resultsEffective associated data for the same type of field in a single due diligence report is limited. Too many recalled results will introduce content from unrelated breeding cycles
similarity threshold0.72–0.80Aquaculture data has high field precision requirements. A threshold that is too low will introduce monitoring data from non-corresponding breeding ponds. A threshold that is too high may miss valid associated records
PARSE_FILE_TIMEOUT_SECONDS300 secondsSingle due diligence reports may include original monitoring logs for multiple batches. Parsing takes a long time, so extending the timeout prevents parsing failures
segment lengthSplit by breeding batch, single segment no more than 1500 charactersAvoid mixing water quality and feeding data from different breeding cycles into the same segment, ensuring semantic relevance during retrieval
rerank return countTop 3–4 resultsPrioritize returning core data directly associated with the target breeding pond for the due diligence, reducing redundant information for subsequent processing

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. Specific issues require individual analysis. It is recommended to test on your own samples before finalizing values.

Three Common Errors

  • Symptom: An error stating localStorage is not defined occurs when executing localStorage.getItem('userId') in a workflow. Cause: The FastGPT workflow runs in a backend service environment, which does not expose browser-side localStorage objects. Directly calling browser APIs causes the error.
  • Symptom: Non-target breeding pond due diligence report content is returned during multi-document knowledge base retrieval. Cause: Precise filtering by document metadata such as document tags or uploaded file names is not configured. Retrieval relies only on global similarity matching.
  • Symptom: Knowledge base retrieval response time exceeds 10 seconds. Cause: recall count and similarity threshold are not adjusted, resulting in too many irrelevant historical breeding data being recalled. This increases indexing and sorting time.

How to Verify Correct Configuration

  • Upload a single aquaculture due diligence report, view the parsed segment list in the knowledge base, and confirm that each segment corresponds to complete field data for a single breeding batch.
  • Enter a specific breeding pond number as a retrieval term, verify that the recalled results only include document fragments associated with that number.
  • Initiate 5 consecutive retrievals, check whether each response time is stable within a reasonable range.
  • Enter the knowledge base configuration page, confirm that rules for recall filtering by document tags or uploaded file names have been set. The open-source version V4.8.22 can use workflow configuration to replace the native AI configuration logic.

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.