What the data for this category looks like
Data sources for infrastructure construction document and material drafting include project bidding documents, construction logs, engineering contracts, acceptance reports, design handover records, and compliance documents. Data updates follow project milestones. Update bidding and contract files when the project starts. Update construction logs and inspection records weekly during the construction phase. Generate sub-acceptance and completion reports during the acceptance phase. Archive all documents after the project is completed. Single documents have significant structural differences. Bidding documents include project overview, qualification requirements, and quotation lists. Each construction log entry includes process name, operating personnel, inspection data, and rectification items. Acceptance reports include sub-acceptance items, qualification judgments, and signature columns. Fields include project number, process ID, acceptance date, compliance standard number, and engineering quantity units. Units include meters, square meters, working days, and ten thousand yuan.
How These Characteristics Create Constraints for Knowledge Base Retrieval and Recall
Dispersed data sources with inconsistent structures require retrieval systems to support both structured field matching and non-semantic text matching. This avoids missing key compliance or technical documents. Document updates are triggered by project milestones, not fixed cycles. Therefore, knowledge base incremental synchronization must tie to project start, construction, acceptance, and other milestones. Do not use scheduled full updates. Single documents vary widely in length. Long documents may include dozens of pages of bidding files. Short records are only hundreds of words of construction logs. This requires adapting chunking and retrieval logic for different text lengths. Professional terms are dense and have synonymous expressions, such as "subgrade compaction degree" and "soil density". This easily causes semantic matching deviations in initial retrieval. An additional semantic correction mechanism is needed.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Professional terms are dense in infrastructure construction documents. This range filters low-semantically relevant results and adapts to matching accuracy for professional text |
limit | Top 8–12 results | Infrastructure documents have long content. Too many retrieved results will exceed the large model context window. 8–12 results cover core reference content |
searchMode | Hybrid search | Infrastructure data includes structured fields (such as project number) and unstructured text. Hybrid search balances semantic matching and precise field matching |
usingReRank | Enabled | Semantic similarity of professional text is easily disturbed by terms. Reranking corrects sorting deviations in initial retrieval and improves the priority of highly relevant documents |
datasetSearchUsingExtensionQuery | Enabled | Professional expressions in infrastructure documents are relatively fixed. Extended queries supplement synonyms and synonymous expressions to improve retrieval coverage |
collectionIds | Fill in the knowledge base ID bound to the current project | Documents of different infrastructure projects are archived independently. Specifying the corresponding collection avoids interference from cross-project irrelevant data on retrieval results |
Three Common Misconfigurations
- Phenomenon: The number of results returned by knowledge base search is always fixed at 5 and cannot be adjusted. Cause: The
limitparameter is not configured correctly, or the corresponding configuration item is not unlocked in advanced settings. - Phenomenon: An error is prompted when importing knowledge base data exported from version v4.9.2 to version v4.12.1. Cause: The field order and required items are not adjusted according to the new version's CSV template, and the project number field of infrastructure documents is not correctly mapped to the target column.
- Phenomenon: The reranking model works normally in the knowledge base search test, but does not take effect during large model calls. Cause: The
usingReRankconfiguration is only enabled in the search test link, and the function is not enabled in the global call settings.
How to Verify Proper Configuration
- Enter the knowledge base search test interface of FastGPT v4.17.0, enter infrastructure professional terms (such as "CFG pile construction requirements"), and check whether the number of returned results matches the configured
limitvalue. - Check the similarity score of a single retrieval result, and confirm that the score falls within the configured
similarityrange. - Trigger a generation task based on infrastructure documents, and check whether the returned content includes highly relevant document fragments after reranking, which conforms to the expected sorting logic.
- Check the retrieval log, and confirm that only data from the
collectionIdscollection corresponding to the current project is called, and no irrelevant documents from other projects are included.
The values provided are common starting points and should be measured against your 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.