What data for this category looks like
Data sources for water treatment intelligent due diligence reports include water quality test documents publicly released by environmental monitoring agencies, daily operation records of water plants, third-party compliance test reports, and pipe network operation logs. Update frequencies fall into three categories:
- Daily operation data is synced and updated daily
- Quarterly compliance test reports are updated every quarter
- Special due diligence documents are updated dynamically as projects progress
Document structure includes structured fields and unstructured attachments. Structured fields cover project number, treatment process type, average daily treatment scale, key water quality indicator test values, and more. Unstructured attachments include scanned equipment maintenance records, rectification notices, process flow diagrams, and more. Most units follow industry standard conventions: average daily treatment scale uses m³/d, and water quality indicators such as COD and ammonia nitrogen use mg/L.
What constraints these characteristics impose on knowledge base retrieval and recall
The above data characteristics create multiple constraints for the knowledge base retrieval and recall link. First, the data includes structured numerical values and unstructured long text. Mixed matching logic for semantic retrieval and full-text retrieval is required. Second, update frequencies vary widely. Daily operation data needs incremental indexing, while quarterly reports need full reindexing. Flexible trigger rules must be configured. Third, the water treatment field has a large number of professional terms and fixed units. The embedding model must adapt to semantic understanding of professional text to avoid matching deviations between values and units. Fourth, some documents are lengthy. A reasonable chunking strategy is needed to retain contextual integrity of process descriptions.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Water treatment due diligence reports often include long documents such as complete operation ledgers and high-definition test reports. 600 seconds covers most parsing scenarios |
embedding_model | bge-large-zh-v1.5 | Adapts to Chinese professional terms in the water treatment field, and can accurately understand professional semantics such as processes and indicators |
chunk_size | 800–1200 characters | Professional paragraphs in water treatment are relatively long. This range retains contextual integrity of process descriptions and avoids logical fragmentation |
recall_top_k | Top 10 results | Due diligence reports need to cover multi-dimensional data including processes, tests, and risks. 10 results balance comprehensiveness and relevance |
rerank_top_k | Top 3 results | Initial recall results may contain weakly relevant content. The top 3 most matching results are retained for context splicing after reranking |
enable_incremental_index | Enabled | Daily operation data is updated incrementally every day. Incremental indexing reduces time and resource consumption of full reindexing |
The parameter values provided on this page are all conventional recommendations used to determine a starting point for configurations. Actual values are affected by material form, data volume, and business rules. Specific issues require specific analysis, and it is recommended to test on your own samples before finalizing settings.
Three common configuration errors
Phenomenon 1
After uploading a water treatment due diligence report in an offline environment, the interface upload status keeps spinning without response, and the console returns 443 port connection timeout logs.
Cause 1
FastGPT's document parsing module relies on public image resources. Offline environments cannot pull dependencies normally, causing the parsing process to block.
Phenomenon 2
When using models such as bge-m3 or bge-large-zh for semantic retrieval, document fragments containing water quality indicators such as COD and ammonia nitrogen cannot be recalled, but full-text retrieval can normally match keywords.
Cause 2
Semantic retrieval field weights are not configured for professional indicators and units in the water treatment field, leading to weakened semantic matching of numerical information.
Phenomenon 3
A large number of non-due-diligence-related equipment promotion documents are mixed in the knowledge base recall results, and the proportion of valid results is very low.
Cause 3
Document tag filtering rules are not configured, and non-due-diligence files are not excluded from the recall scope.
How to confirm configurations are correctly set
- Upload a typical water treatment due diligence report, check the parsed chunking results, and confirm that the chunk length matches the preset
chunk_sizerange. - Enter professional search terms such as "MBR process operation records" and "COD exceeding standard rectification", compare the recall results of semantic retrieval and full-text retrieval, and confirm that semantic retrieval can match contextual associations of professional terms.
- Check the index update logs in the console, and confirm that the trigger logic for incremental updates or full reindexing matches the preset update frequency requirements.
- Test the file upload and parsing process in an offline environment, confirm that there are no external dependency pull requests, and avoid blocking issues.
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.