This document applies to FastGPT v4.17.0.
What This Category's Data Entails
The data sources for this category include material grade standards, customer technical requirements, temperature control instrument raw data, furnace temperature uniformity test reports, hardness tester and metallographic equipment test results, and workshop MES system records. Data is updated per furnace run or batch. Single batch associated documents include heat treatment process cards, charging charts, furnace temperature curves, hardness and metallurgical reports, equipment inspection forms, and exception handling documents, among others. Document fields include heating temperature (unit ℃), holding time (unit h or min), cooling medium, hardness value (unit HV or HRC), furnace temperature deviation value, and others. Some documents include structured fragments of time-temperature correspondence curves.
Constraints Imposed on Vector Models and Indexing
This category of data is updated per furnace run or batch. A single process Q&A requires linking multiple documents from the same batch. The index must therefore support precise recall by batch dimension. Documents contain large numbers of numerical process parameters and structured curve fragments. The vector model must adapt to semantic matching of numerical fields, and avoid breaking the temperature-time correspondence during chunking. Some data comes from raw device instrument logs, with formats differing from standardized process cards. The index must support embedding processing for multi-format text. Minor deviations in process parameters directly affect the mechanical properties of parts. The accuracy requirement for retrieval results is higher than that of general scenarios.
Configuration Settings
| Configuration Item | Recommended Values | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapt to the single-section text length of heat treatment process cards and furnace temperature curves, avoid splitting that breaks the temperature-time correspondence |
| Overlap Length | 100–150 characters | Retain process connection information between adjacent furnace runs or batches, prevent chunking from cutting off charging position descriptions |
| Similarity Threshold | 0.75–0.85 | Distinguish minor differences in process parameters, avoid mismatching heat treatment regimes of different material grades |
defaultToken | 1024 | Control the number of tokens for a single embedding segment, adapt to the single-section text length of heat treatment documents |
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. |
| Single Search Limit | Top 6 entries | Cover associated documents for a single furnace run, such as process cards, furnace temperature curves, and hardness reports |
Three Common Misconfigurations
- Retrieval results return multiple unrelated bearing heat treatment processes. The cause is a similarity threshold set too low, which mismatches workpiece processes of different material grades with similar temperature parameters. This may lead to subsequent errors in process execution.
- Timeout errors occur during retrieval, with error codes showing request duration exceeding the threshold. The cause is a single search limit set too high, which loads too much cross-batch furnace temperature curve data at once.
- Embedded vectors fail to match the holding time parameters in process cards. The cause is a chunk length set too short, which cuts off the temperature-time corresponding relationship during chunking, leading to loss of semantic information.
How to Verify Proper Configuration
- Upload a single heat treatment process card, run a retrieval test, and check if the returned result chunks retain complete heating-holding-cooling process descriptions.
- Enter a query containing specific temperature and time values, verify that the similarity scores of the returned results fall within the preset 0.75–0.85 range.
- Simulate an exception query such as abnormal hardness for a specific furnace run, confirm that the retrieval results include the corresponding exception handling document and metallurgical report.
- View the embedding task logs, confirm that no token limit exceeded errors occur, and that the
defaultTokenandmaxTokenconfigurations are followed.
The values provided are common starting points, and should be evaluated against the reader's own samples.