What the data for this category looks like
Infrastructure construction policy and regulatory basis data comes primarily from official policy documents of the Ministry of Housing and Urban-Rural Development, provincial housing and urban-rural development authorities, local government-issued infrastructure-related regulatory texts, and industry standard specifications. Updates are synchronized with new policy releases, with no fixed cycle. Most documents are PDFs or structured policy clauses. Fields include policy document number, issuing authority, release date, applicable scenario, clause content, and applicable region. Identifier formats follow document number conventions, and dates use the YYYY-MM-DD format. Information covers engineering categories, applicable section scope, and similar details.
What constraints these characteristics impose on vector models and indexes
Infrastructure construction policy texts are long and structured. When chunking these texts, preserve complete clause logic. Do not force truncation that causes semantic breaks. Chunk length must match policy clause length to avoid splitting policy logic. Include metadata fields including document number, release date, and applicable region in index metadata for subsequent precise recall filtering. Since updates have no fixed cycle, use incremental indexing adapted to non-fixed update rhythms to avoid excessive full indexing time. Vector recall must support filtering by metadata such as release date and applicable region. This matches policy applicable scope and improves recall accuracy.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 600–1000 characters | Matches the length of infrastructure construction policy clauses, preserves single policy logic, avoids semantic breaks |
| Overlap Length | 100–200 characters | Maintains contextual coherence between adjacent policy clauses, avoids logical gaps across clauses |
| defaultToken | 8191 | Adapts to the maximum token limit of text-embedding-ada-002, aligns with general vector model standards |
| maxToken | 4096 | Matches the context length requirements of infrastructure construction policy Q&A, balances semantic completeness |
| Similarity Threshold | 0.72–0.85 | Filters low-relevance policy recall results, avoids interference from irrelevant clauses |
| Single Search Limit | Top 10 entries | Controls the number of recalled policy clauses, balances recall comprehensiveness and query efficiency |
Three common mistakes
- Phenomenon: After the knowledge base completes full indexing, search tests take too long. Logs show single query time exceeds 5 seconds. Cause: No single search limit is set. Too many policy texts are recalled, and the vector database scan range is too large. This increases subsequent processing pressure.
- Phenomenon: After configuring text-embedding-ada-002 and adding it to the OneAPI channel, an error "No matching embedding results found" pops up when calling the knowledge base. Cause: The correct model identifier is not specified in FastGPT's vector model configuration, or the model name in the configuration file does not match the name in the OneAPI channel.
- Phenomenon: Calls to the embedding-2 model fail. Logs show the model was not found. Cause: The custom-deployed embedding-2 model is not synchronized to FastGPT's available model list, or the model is not enabled in the channel configuration.
How to confirm the configuration is complete
- Log in to the FastGPT v4.17.0 knowledge base management module. Check whether the number of indexed policy data entries matches the actual imported number of entries.
- Enter a known policy keyword and initiate a search. Check whether the similarity scores of returned results fall within the configured threshold interval.
- Open the vector model configuration page. Confirm that the added embedding model and deployment channel are correct, and that the
defaultTokenandmaxTokenparameters in the configuration file meet the configuration requirements. - Check system logs. Confirm that no timeout errors appear in vector recall query times.
The values provided here are common starting points. They should be evaluated 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.