What this category of data includes
This category of data originates from wafer test maps, manufacturing execution systems (MES), bonding and molding equipment, ATE test machines, X-ray/SAM inspection equipment, and customer specification documents. Daily production and maintenance documents cover packaging process cards, bonding and molding parameter tables, incoming wafer maps, test program version lists, yield reports, and failure analysis reports. Data updates occur per production batch, after each chip test, and daily. It spans all process links from incoming wafer materials to finished product testing.
Constraints on multi-turn dialogue and prompt engineering
This category of data has high update frequency and fine granularity. Multi-turn dialogue must retain core context identifiers such as batch ID and incoming wafer map number, to avoid mixing production data across batches. There are many specialized daily document types. Prompt engineering must clearly define retrieval scopes, such as only retrieving bonding machine parameters or failure analysis report documents, to prevent cross-document field confusion. Equipment data comes from multiple hardware and system sources. Prompt engineering must specify data source tags, to ensure called equipment parameters match the model currently being inspected and maintained. Redundant single-chip test details must be filtered out, and only batch-level summary information is retained for inspection and maintenance Q&A.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Maximum context length | 800–1200 characters | Matches the average length of single-batch production documents for this category, avoids loading redundant single-chip test data |
| Variable binding | Bind chatId and batch ID | Ensures same-batch dialogue contexts are not disturbed by cross-batch calls, adapts to the variable configuration logic of FastGPT v4.17.0 |
| Maximum retained message history | First 10 dialogue turns | Filters redundant single-chip test details, retains only key dialogue nodes for batch-level inspection and maintenance |
| Retrieved document scope | Only retrieve packaging process cards, bonding and molding parameter documents | Matches the daily document types for this category, prevents calling unrelated wafer manufacturing documents |
| Timeout threshold | 60 seconds | Adapts to the average response time for retrieving equipment data from MES, avoids dialogue interruptions |
Three common configuration errors
- Request Time errors occur during dialogue, with empty return results. Cause: No timeout threshold configured for this category’s equipment data retrieval. Response times exceed default limits when pulling equipment data from multiple sources including MES and ATE test machines.
- Tool call results only appear after re-entering the dialogue, and are not displayed synchronously during the dialogue. Cause:
variablesand the current batch’s equipment ID are not correctly bound, so the dialogue context cannot link to tool call return data. - The interface displays in English after iframe embedding, and cannot be switched to Chinese. Cause: No language parameter is set in the iframe embed code, or the parameter value does not follow FastGPT’s interface language configuration rules.
How to verify correct configuration
- Initiate an inspection query for the current production batch’s equipment, confirm the dialogue context includes the bound batch identifier and
chatId. - Call the tool to retrieve equipment parameters, verify returned results match the current batch’s bonding or molding parameters, with no unrelated document content included.
- Start consecutive multi-turn maintenance Q&A dialogues, check for timeout errors, adjust the timeout threshold to match the actual response time of equipment data retrieval.
- Validate the interface language in the iframe embedding scenario, confirm configuration parameters match the required business language type.
The values provided are common starting points and should be evaluated against the reader’s own samples.