What the data for this category looks like
Data sources for gas intelligent due diligence reports include internal pipe network ledgers from gas operation enterprise management systems, safety inspection reports from third-party testing institutions, and compliance disclosure documents from municipal public utility supervision platforms. Update rhythms vary across sources: basic pipe network parameters are updated every six months, single inspection records are updated in real time per inspection batch, and annual compliance documents are updated per regulatory inspection cycle. Documents take structured tables as their core, supplemented by unstructured hazard descriptions and on-site photo annotations. Fields include pipe inner diameter, detection pressure value, inspection cycle, and longitude and latitude coordinates. Each field has clear physical units attached.
Constraints on vector models and indexing
Structured numerical and geospatial fields account for a large share of gas due diligence reports. General vector models cannot accurately encode physical features such as longitude and latitude and pressure values. Specialized vector models that support structured data encoding are required. Large differences in update rhythms across data sources mean full index rebuilding consumes significant computing resources. Support for batched incremental indexing is needed. Fields in documents are closely linked. For example, a single section may include both pipe location and corresponding pressure parameters. Contextual association must be retained during chunked indexing to avoid retrieval accuracy degradation from feature fragmentation.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
embedding_model | bge-m3 or text-embedding-v3 | Both models support encoding of structured numerical and geospatial features, and are compatible with the multi-type fields in gas due diligence reports |
chunk_size | 800–1200 characters | Gas due diligence reports contain content with linked multiple fields. This length retains field associations within a single chunk and avoids feature splitting |
enable_spatial_index | Enabled | The reports include longitude and latitude coordinate fields. Spatial indexing optimizes the recall efficiency and accuracy of geospatial features |
index_refresh_interval | 1 hour (inspection data) or 7 days (basic pipe network data) | Matches the update rhythms of different data sources and avoids resource consumption from full index rebuilding |
similarity_threshold | 0.75–0.85 | The semantic similarity of gas safety-related features has high discriminability. This range filters low-relevance recall results |
top_k | Top 8–12 results | Due diligence reports have many feature dimensions. An appropriate number of recall results covers all relevant information |
The parameter values provided on this page are common recommended starting points for configuration. Actual values are affected by material form, data volume, and business rules. Specific issues require case-by-case analysis. It is recommended to test on your own samples before finalizing settings.
Three common errors
- Phenomenon: The interface displays the prompt "No available vector model channel". Cause: The API key and access address of the corresponding vector model have not been configured in model vendor management, and the channel has not been bound to the default group.
- Phenomenon: The number of vector recall results is less than the configured
top_kvalue. Cause: The configuredsimilarity_thresholdis too high, resulting in fewer recall results meeting the threshold than the specified number. - Phenomenon: Newly added inspection data is not synchronized to the index library. Cause: The
index_refresh_intervalparameter has not been adjusted to match the update rhythm of inspection data, and an overly long refresh cycle is used.
How to confirm the configuration is complete
- Enter the model vendor management page, check the access status of the vector model, and confirm that it shows "Connected".
- Upload a gas due diligence report to trigger vector indexing, and check whether the indexing logs contain vector encoding records for fields such as longitude and latitude and pipe pressure.
- Initiate a retrieval request related to gas safety or pipe network parameters, and check whether the similarity scores of the recall results fall within the configured
similarity_thresholdrange. - Modify the
index_refresh_intervalparameter, wait for the corresponding cycle, and check whether newly added inspection data has been synchronized to the index library.
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.