What the Data for This Category Looks Like
New energy development service inquiry data sources include project approval documents, grid connection process guides, environmental impact assessment reports, local energy policy documents, and on-site operation and maintenance inquiry records. Update frequency varies by content type:
- Policy documents are updated quarterly or annually
- Project files are updated in real time as projects progress
- Operation and maintenance records are synced daily
Document formats include scanned PDFs, structured approval forms, and Word process manuals. Fields include project number, approval node, compliance requirements, installed capacity, and more. Units use professional energy units such as megawatts (MW), kilowatt-hours (kWh), and hectares (hm²).
Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
Scanned PDF documents require OCR recognition to extract valid text, which adds configuration requirements for the preprocessing link. Fixed fields and numerical units in structured approval forms require retrieval to balance semantic matching and field accuracy. This avoids recall results with unit confusion. Multi-source, multi-format data must be managed by project classification. Otherwise, the retrieval scope may mix irrelevant content across projects. Chunking processing for long documents such as environmental impact assessment reports must retain context association. Otherwise, semantic fragmentation will occur and reduce recall accuracy.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Aligns with FastGPT v4.17.0 parameter logic, filters irrelevant recall from general energy content |
limit | Top 8 entries | Covers inquiry needs across new energy project approval, grid connection, operation and maintenance, and other links. Avoids excessive results that create screening burden |
searchMode | Hybrid retrieval | Balances field matching for structured forms and semantic recall for unstructured policy text |
usingReRank | Enabled | Resolves semantic scattering after long document chunking, optimizes recall result sorting logic |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands professional terms, such as expanding "photovoltaic" to expressions for relevant sub-scenarios, improving recall coverage |
collectionIds | Bound by project classification | Limits retrieval scope to knowledge base collections associated with the current new energy project, avoiding irrelevant cross-project content |
Three Common Misconfigurations
- Phenomenon: After scanned new energy project PDF documents are uploaded, retrieval does not return matching content. Cause: OCR preprocessing is not enabled, so the system cannot extract text from scanned images.
- Phenomenon: After knowledge base data is added via the business interface, calling the retrieval interface returns empty results. Cause: The newly added knowledge base collection ID is not bound to the
collectionIdsretrieval configuration parameter, so the retrieval scope does not cover the new data. - Phenomenon: When an overly long inquiry question is entered, retrieval result relevance drops significantly. Cause: The ultra-long input exceeds the embedding model's context length limit, leading to semantic encoding distortion and inaccurate document matching.
How to Verify Proper Configuration
- A scanned new energy policy PDF can be uploaded. After system processing completes, an inquiry question containing keywords from the document can be entered, and confirmation made that the corresponding content is recalled.
- Knowledge base data can be added via the business interface. Afterward, the retrieval interface can be called and the corresponding collection ID specified, and confirmation made that returned results include fragments of the new data.
- A relatively long inquiry question can be entered, and confirmation made that retrieval result relevance does not show abnormal fluctuations.
- The reranking function can be enabled, and confirmation made that recall result sorting aligns with expected matching to the inquiry question.
The values provided here are common starting points. Adjust them based on the target deployment's sample data and use cases.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.