What data for this category looks like
This category of data primarily originates from raw material quality certificates, heating furnace systems, press collection systems, MES systems, non-destructive testing instruments, and customer forging drawings. Daily generated documents include forging process specifications, cutting cards, heating furnace records, forging process cards, mold life records, heat treatment work orders, and ultrasonic testing reports. Data is updated per individual forging part, batch of materials, or furnace heat. It includes dedicated fields such as furnace batch number, forging ratio, fiber direction, and detected defect type.
What constraints these characteristics place on vector models and indexes
This category of data has a high update frequency, with new data created per batch or furnace heat. Vector indexes must support incremental updates to avoid the time cost of full reconstruction. Document types include both structured numerical records and unstructured process text, so the system must support vectorization logic for multimodal data. Dedicated fields like furnace batch number and forging ratio must be bound as metadata to their corresponding vectors, to enable associative retrieval by batch later. Exception reports involve risk of forging scrap, so retrieval recall accuracy must align with business verification needs to prevent missing critical defect information.
Configuration settings
The following configurations are built for FastGPT v4.17.0, and adapted to the business requirements of forging and die forging scenarios.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the paragraph length of forging process specifications, avoiding cutting that disrupts the integrity of step descriptions |
| Overlap Length | 100–150 characters | Covers associated information of adjacent steps, preventing process steps from being truncated by chunking |
| Similarity Threshold | 0.75–0.85 | Matches the semantic similarity range of defect descriptions in exception reports, filtering low-relevance retrieval results |
| Single Search Limit | Top 10 entries | Balances retrieval efficiency and recall completeness, covering associated documents for common anomaly types |
defaultToken | 4096 | Adapts to the text length of a single process document and exception report, meeting basic context call requirements |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
Three common configuration mistakes
- Phenomenon: Retrieval results do not return bound ultrasonic testing report images. Cause: Image vectorization configuration is not enabled, and image files are not linked to the corresponding forging part's process documents.
- Phenomenon: The same anomaly query triggers a vector database request every time, creating additional time consumption. Cause: The query caching function is not enabled, or the cache validity period is not set to match the update frequency of business data.
- Phenomenon: No retrievable index is generated after adding data to a collection. Cause: The functions of adding data to a collection and creating training orders are confused, and only text is uploaded without triggering the batch vectorization process.
How to verify correct configuration
- Upload a typical forging process specification and ultrasonic testing report, perform a retrieval test, check the chunk integrity and association matching of the returned results, and confirm that the
maxTokensetting meets the needs of multi-document splicing. - Initiate two identical anomaly queries, compare the response speeds of the two attempts, and confirm that the second query has significantly reduced latency, verifying that the caching function is active.
- View the index list in the vector database, confirm that the number of bound documents matches the number of uploaded documents.
- Try selecting the configured vector model, confirm that it loads normally and generates embedding vectors, with no model selection errors.
The values provided are common starting points and should be measured against samples relevant to the intended use scenario.