What Data for This Category Looks Like
Data for this category comes primarily from daily outputs of automotive, aerospace, and mechanical assembly enterprise tooling departments, as well as dedicated tooling suppliers. Included documents are tooling requirement sheets, positioning datum schemes, fixture drawings, gauge operation manuals, acceptance reports, calibration records, and change histories. Data sources cover product digital models and GD&T, process routes, coordinate measuring machine data, on-site trial feedback, and metrology systems. Updates follow project milestones, each product change, and regular calibration cycles. Documents contain specialized professional fields such as positioning datum parameters, interference verification records, and repeatability and reproducibility data.
Constraints on Vector Models and Indexing
The multi-modal, dense professional parameter data of this category creates multiple constraints for the vector model and indexing workflow. First, documents include multi-modal content such as engineering drawing annotations and metrology values. Select vector models that support professional engineering text and numerical data to avoid losing semantic meaning of critical professional parameters. Second, data updates dynamically per project. Single-project dataset scale fluctuates with product complexity. Indexing must support incremental updates and version association to ensure only currently valid data is called during cross-review. Third, documents include time-sequenced change histories. Indexing must retain version identifiers to distinguish new and old data. Fourth, documents include precise parameters with units such as GD&T tolerances and coordinate measuring machine readings. Vector extraction must retain unit associations to ensure retrieval matching accuracy.
Configuration Settings
The following configurations are adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the average paragraph length of documents like fixture drawing descriptions and calibration records. Prevents splitting critical professional parameters such as positioning datums across chunks. |
| Overlap Length | 100–150 characters | Retains contextual association of positioning and clamping logic across paragraphs. Prevents critical verification logic from being split apart. |
maxToken | 4096 token | Matches the maximum input limit for multi-modal document embedding. Accommodates long paragraphs containing drawing annotation text. |
defaultToken | 512 token | Used for basic embedding of short documents such as single calibration records. Balances computational efficiency and semantic completeness. |
| Similarity Threshold | 0.75–0.85 | Filters low-match irrelevant documents. Ensures only same-type or cross-project tooling data is retrieved during cross-review. |
| Single Search Limit | Top 10 results | Controls the scale of retrieval results. Fits the need for multi-document association verification during tooling review. |
Three Common Configuration Mistakes
- Phenomenon: Indexing takes significantly longer than expected. Single-project dataset indexing exceeds 2 hours. Cause: Chunk parameters are not adjusted based on individual document length. This creates an excessive number of chunks, with unreasonable batch scheduling for vector generation and index writing.
- Phenomenon: Indexing tasks fail to trigger normally in a docker-compose deployment environment. The interface shows empty index task status. Cause: The corresponding data volume is not mounted in the docker-compose configuration file. This prevents the indexing script from reading local document data sources.
- Phenomenon: Retrieval results include large volumes of old version change records. This leads to incorrect compliance judgments for current tooling during cross-review. Cause: Document version fields are not associated in indexing configuration. Vector similarity cannot distinguish between new and old data, resulting in invalid historical data being included.
How to Verify Correct Configuration
- Upload a single typical tooling requirement sheet. Run a vector generation task. Check that chunking results retain complete positioning datum parameters, with no splitting of critical parameters.
- Initiate a multi-document cross-retrieval. Confirm the number of returned results matches the configured single search limit, and that result matching degrees fall within the preset similarity range.
- Review indexing logs. Confirm incremental update tasks trigger automatically according to project change milestones, with no duplicate indexing or missed documents.
- Test embedding generation for a fixture drawing document with annotations. Confirm no errors occur, and vector generation time meets expected levels.
The values provided are common starting points and should be measured against the reader's own samples.