What data looks like for this category
Data sources for semiconductor packaging and testing production ledgers include wafer test maps, MES systems, bonding and molding equipment logs, ATE tester outputs, X-ray/SAM inspection reports, and customer specification documents. Data updates occur per production batch, single-chip test, and daily equipment operation statistics. Daily ledger documents include packaging process cards, bonding and molding parameter records, wafer incoming material maps, test program version files, yield statistics documents, and failure analysis reports. Core fields include batch identifiers, equipment operation parameter thresholds, test voltage ranges, chip pin parameters, and test result identifiers. No unified standard units exist. Matching measurement rules must align with equipment type and document type.
Constraints for Multi-turn Dialogue and Prompt Engineering
Multi-source heterogeneous data sources require multi-turn dialogue to gradually align field definitions across different documents, to avoid parameter confusion between documents. High-frequency updated production data requires limiting the valid range of dialogue context, to prevent interference from old batch data on current queries. Non-standardized fields across different documents require prompts to clearly define field mapping rules, to ensure parsed results match the exclusive format of packaging and testing ledgers. Error risks in production links require dialogue to verify matching between batch identifiers and parameters, to prevent mixed batches or incorrect test program usage. Continuous association with the context of the current production batch is required during multi-turn interaction, to avoid data confusion across batches.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Maximum Context Length | 8000–12000 token (calibrated via actual testing) | Covers the combined length of multiple ledger documents for a single batch, and meets information retention needs for multi-turn dialogue |
| Single-Turn Dialogue Token Limit | 4000 token | Restricts the number of parameters in a single query, to avoid exceeding the field parsing boundary of packaging and testing ledgers |
variables Initialization | Bind three core variables: batch number, equipment ID, and test program version | Matches the core retrieval dimension of batch-based ledger management for packaging and testing, and simplifies parameter transfer for multi-turn dialogue |
messages Truncation Strategy | Retain the most recent 20 dialogue turns | Filters invalid context from old batches, and focuses on dialogue information for the current production link |
chatId Association Configuration | Bind the unique identifier of the current production batch | Adapts to FastGPT v4.17.0's dialogue context management rules, ensuring that dialogue context only associates with ledger data for the corresponding batch per turn |
Three Common Configuration Mistakes
- Phenomenon: Uploaded wafer incoming material maps and packaging process cards cannot be read during dialogue, while the knowledge base file upload function operates normally. Cause: Context association for file parsing was not enabled in multi-turn dialogue configuration; only a static knowledge base index was configured.
- Phenomenon: Generated SQL query result counts are abnormal, and cannot match equipment parameter fields from ATE testers. Cause: The prompt did not clearly define exclusive field mapping for packaging and testing ledgers, and did not restrict retrieval variables to batch number and equipment ID.
- Phenomenon: After switching production batches in multi-turn dialogue, ledger data from old batches remains in the context. Cause: The current production batch was not bound via
chatId, andvariableswere not reset before each dialogue turn.
How to Verify Proper Configuration
- Initiate a query that includes a production batch number and bonding pressure parameter, and confirm returned results only associate with ledger data for the current batch.
- Upload a single packaging process card and initiate a dialogue, and confirm parsed fields match the preset
variablesvariables. - Initiate three consecutive queries with different parameters, and confirm the context only retains valid information from the current dialogue turns.
- Trigger a SQL generation action, and confirm generated statements follow the exclusive field naming rules of packaging and testing ledgers.
The values provided are common starting points, and should be evaluated against the reader's own samples.