What data looks like in this scenario
Data for shipbuilding internal office and employee services scenarios comes from product data management (PDM) systems, enterprise resource planning (ERP) systems, office automation (OA) systems, equipment operation logs, employee attendance records, and job operation manuals. Data types include structured production work orders, semi-structured ship design documents, unstructured operation logs, and employee training materials. Update frequency varies widely by business scenario: production work orders update in real time per production shift, design drawings update in phases per project milestones, and employee handbook-style documents update quarterly. Fields include professional identifiers such as ship section numbers, weld inspection report numbers, and main engine models. Units include millimeters, tons, operating hours, and other common industrial units.
What constraints these characteristics impose on the vector models and indexing link
Multi-field attributes of structured production work orders require retaining field association information during indexing to avoid chunking breaking business logic. Semi-structured design documents contain long paragraphs, which impose clear requirements for chunk length and maxToken settings. Real-time updated production work orders need incremental indexing support to reduce resource consumption from full reconstruction. Unstructured operation logs have wide variations in text length, requiring flexible chunking rules to adapt to different content lengths. Uniqueness of professional fields such as ship section numbers requires association constraints during indexing to prevent duplicate data from being stored.
How to set configurations
Reference only; verify against the target FastGPT release and deployment documentation.
| Config Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the semi-structured paragraph length of ship design documents and avoids splitting technical terms |
| Overlap Length | 50–80 characters | Retains contextual association between chunks and prevents professional information in long paragraphs from being truncated |
| Similarity Threshold | 0.72–0.85 | Matches semantic similarity of technical terms in shipbuilding scenarios and filters low-relevance results |
| Single Search Limit | Top 10 results | Controls the number of returned results to meet the precise retrieval needs of internal office work |
maxToken | 4096 token | Adapts to the input length limit of mainstream vector models and meets the needs of long document chunking |
defaultToken | 256 token | Processes vector generation for short text such as employee work order summaries and avoids redundant calculations |
Three common mistakes
- The total volume of data after vectorization is thousands of entries fewer than the source files. Cause: Empty fields or invalid rows were not filtered during chunking. Shipbuilding work order data contains unfilled equipment number fields, which are automatically skipped.
- The collection is created successfully but the index never finishes building on the page. Cause: Incremental indexing trigger rules were not configured, real-time updated production work order data was not synchronized to the vector database, or index shard configuration was insufficient causing timeouts.
- Vector index results work normally locally, but vector scores are consistent when running via Docker image. Cause: The correct vector model cache directory was not mounted inside the Docker container, causing the model to be re-pulled on each call, or environment variable configuration anomalies caused changes to the model call logic.
How to confirm proper configuration
- View vector task logs to confirm that the number of chunked documents matches the number of valid rows in the source data.
- Retrieve a known technical term and verify that the similarity distribution of returned results falls within the preset threshold interval.
- Upload a test fragment of a ship design document to verify that index generation time matches the expected configuration.
- Check incremental indexing trigger records to confirm that real-time updated work order data has been synchronized to the vector database.
The values provided 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.