What the data for this category looks like
The data mainly comes from official policy documents, industry standard texts, and compliance guidelines issued by the National Development and Reform Commission, the National Energy Administration, and local energy authorities. Updates have no fixed cycle, and are triggered by the release of new policies or revisions of industry norms. A single document usually includes a structure with document number, issuing authority, release date, clause text, supporting attachments, and other elements. Fields cover policy hierarchy, applicable regions, implementation cycle, and more. Some attachments include quantitative indicators, with units such as yuan per kilowatt-hour, ten thousand kilowatts of installed capacity, and others.
What constraints these characteristics impose on the vector models and indexes link
Official policy documents mostly have long text structures and include attachments with varying formats, which places requirements on the coherence of text splitting and structure retention. Updates have no fixed cycle, so flexible configuration supporting incremental indexing and regular full updates is needed. Differences in policy hierarchy and applicable regions require the indexing link to distinguish the priority and coverage of different policies, avoiding confusion between cross-level policies. Some documents include quantitative indicators, so vector matching needs to balance clause semantics and indicator accuracy, preventing compliance risks caused by fuzzy matching.
How to set the configurations
| Configuration Item | Recommended Approach | Basis for This Approach |
|---|---|---|
| Chunk Length | 800–1200 characters | New energy policy documents mostly consist of long text clauses; this range can preserve the integrity of a single policy logic and avoid semantic fragmentation |
| Overlap Length | 100–200 characters | Compensates for semantic gaps after long text splitting, ensuring coherence of cross-chunk policy clauses |
| Similarity Threshold | 0.75–0.85 | Policy compliance Q&A requires precise clause matching; this range can filter low-correlation results and avoid misuse of non-applicable policies |
| Single Search Limit | Top 8 results | Covers multi-clause association scenarios of new energy policies, while controlling context length to avoid overflow |
maxToken | 4000 token | Matches the context length limit built into FastGPT v4.17.0, adapting to the multi-clause integration needs of policy Q&A |
| Reranking | Enabled | Optimizes the ranking logic of vector recall results, prioritizing policy clauses directly related to the query |
Three common mistakes
- Phenomenon: The locally used vector model is inconsistent with the server-side one, and retrieval results do not meet expectations after uploading documents. Cause: Embedding dimensions and semantic mapping logic of different vector models differ, and ununified configurations lead to incompatible vector spaces.
- Phenomenon: A single policy file continuously shows the "indexing" status with no progress updates or error prompts. Cause: The file contains complex nested tables or oversized attachments, exceeding the default processing capacity of chunk processing, causing the indexing process to block.
- Phenomenon: When multiple knowledge bases are retrieved together, the policy content of a specified knowledge base cannot be called preferentially. Cause: No knowledge base priority rules are configured, or the rules are not bound to the current Q&A scenario, leading to random calls of results from each knowledge base during retrieval.
How to confirm the configuration is properly set
- A single typical policy document is uploaded, the indexing progress log is checked, and the number of chunks is confirmed to fall within the expected range.
- A targeted policy question is input, the similarity scores and ranking of the recall results are verified, and the threshold and search limit are adjusted to meet business requirements.
- Two test knowledge bases are imported, priority rules are configured, a Q&A is initiated, and it is verified that the results preferentially match the content of the preset high-priority knowledge base.
- The vector model configuration items are checked, the model parameters used locally and on the server side are confirmed to be consistent, to avoid embedding space mismatch.
The values provided here are common starting points and should be measured against the reader's own samples.
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.