What this category of data looks like
Survey and design business data is primarily sourced from full-cycle project documents including survey reports, site reconnaissance records, design plan descriptions, business process approval forms, and accompanying drawing OCR text. Data updates are triggered by project milestones: initial documents are uploaded when a project starts, modified as needed during the approval phase, and archived after project completion. Documents include structured fields such as project number, survey depth, and approval time limit, alongside unstructured text paragraphs. Single document lengths vary widely, ranging from a few hundred characters of milestone descriptions to tens of thousands of characters of complete survey reports.
What constraints these characteristics impose on vector models and indexing
The mixed structure of survey and design data requires indexes to support both semantic recall and metadata filtering, to accurately locate survey points or approval processes for the corresponding project. Document updates are triggered by project milestones, so indexing logic must adapt to on-demand incremental index construction to avoid wasting resources on invalid indexes. Some text from drawing OCR recognition contains formatting noise, so basic line break and redundant symbol cleaning must be performed before indexing. The wide range of single document lengths requires support for dynamically adjusting chunking parameters to adapt to text fragments of different lengths.
How to set configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the common paragraph length of survey reports and approval descriptions, avoiding overloading single chunks |
| Overlap Length | 100–150 characters | Preserves cross-chunk process context association, preventing key approval nodes from being truncated |
| Similarity Threshold | 0.72–0.85 | Calibrated based on actual measurements, filters low-relevant survey points or process descriptions |
| Single Search Limit | Top 8 entries | Matches the common number of associated documents for survey and design projects, avoiding exceeding context quotas |
maxToken | 16000 | Adapts to the standard context window of large models, accommodating multiple recalled survey and approval documents |
defaultToken | 2000 | Configures the basic token quota for single-round question answering, supporting accurate recounting of complete process steps |
Three common misconfigurations
- Symptom: The interface does not allow selection of an index model, and a pop-up prompts parameter exceptions. Cause: Pre-binding of the vector model has not been completed, or the format of uploaded survey and design documents does not meet index parsing requirements, causing the system to fail to load available index models.
- Symptom: Source document information for the corresponding text block is not returned when answering process questions. Cause: The metadata extraction switch was not enabled in the index configuration, and fields such as document name and project number were not included in the index metadata binding scope.
- Symptom: A large number of irrelevant cross-project document texts appear in recall results. Cause: No project number was configured as a metadata filtering condition, resulting in index recall that does not limit the scope of documents for the current business scenario.
How to confirm correct configuration
- A single survey report is uploaded, a query for a specified survey point within the report is initiated, and recalled results are checked to confirm they include the corresponding text block from the report.
- The metadata list in the index configuration interface is reviewed, and survey and design-specific fields such as project number and document name are confirmed to be selected as filtering conditions.
- A test query is initiated, and the similarity scores of returned results are checked to confirm they fall within the configured threshold range.
- Source tracing information in question-and-answer results is reviewed, and corresponding document names and paragraph positions are confirmed to be displayed.
The values listed above are common starting points and should be evaluated against local 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.