What the Data for This Category Looks Like
The data for this category typically comes from silicon wafer incoming material records, workshop MES system logs, operation logs of diffusion furnaces, coating equipment, and printing presses, EL and IV test reports, and laboratory quality inspection data. Daily produced documents include process cards, diffusion and coating parameter tables, printing and sintering records, EL detection images, IV grading data, and yield analysis reports. Data is updated daily in sync with production batches and single-chip test results. Each document contains fields such as batch number, process parameter values, detection time, and electrical performance indicators, with clear physical quantity units attached.
What Constraints These Characteristics Impose on the Knowledge Base Retrieval and Recall Link
This category has high-frequency, fine-grained data updates. A single retrieval must cover the latest batch-level and single-chip-level data, which imposes clear requirements for retrieval real-time performance. Multiple types of documents coexist, including structured parameter tables, visual detection images, and text reports, which requires the retrieval system to support content parsing across different formats. Process links and detection indicators are tightly bound, so retrieval must accurately match specific fields such as process windows and electrical performance grading to avoid interference from irrelevant results. At the same time, the accuracy of retrieval results directly affects process adjustment and quality inspection judgment, so the proportion of false recalls and missed recalls must be strictly controlled.
How to Configure the Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Meets the accuracy requirements for matching process parameters and detection indicators, and avoids low-relevance results from being included in the retrieval set |
limit | Top 12–18 results | Covers multiple sets of process and detection data involved in a single batch production, while controlling the context length |
searchMode | mixedRecall | Balances exact matching of structured parameters and semantic matching of unstructured documents, adapting to multi-type data |
usingReRank | Enabled | Performs secondary ranking on initial retrieval results to filter redundant associated results in EL detection images and IV data |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval terms to cover aliases of process parameters and industry-standard expressions, such as matching "方阻" to "sheet resistance" |
collectionIds | Knowledge base collections divided by production department | Limits the retrieval scope to documents from corresponding departments such as process, equipment, and quality inspection, narrowing the scope of ineffective retrieval |
Three Common Mistakes
- Symptom: Plugin confirmation for FastGPT v4.17.0 beta4 fails with a MongoServerError. This error alone does not establish a compatibility conflict between the plugin and MongoDB 4.4.29; inspect the complete logs, the database request issued by the plugin, and account permissions. The v4.17.2 deployment config documents 4.4.29 as a fallback for CPUs without AVX (see the official deployment configuration).
- Phenomenon: After importing a CSV-format knowledge base between FastGPT instances of the same version, the hierarchical distinction of multiple files cannot be retained. Cause: The export and import process does not correctly recognize the associated metadata of multiple files, only flattening the content into a single-file format.
- Phenomenon: When the AI retrieves table data, only some entries are returned, and all 50 items in the table are not covered. Cause: The initial retrieval
limitparameter is set too low, or the exact matching mode for structured data is not enabled, causing some low-similarity entries to be filtered out.
How to Confirm the Configuration Is Correct
- Run a retrieval test for single-batch process parameters, and verify that the returned results include all associated documents and fields of the batch.
- Import a backup of the knowledge base of the same version, and confirm that the hierarchical structure of multiple files is fully retained after import.
- Run a plugin compatibility check, and confirm that there are no syntax conflicts between the current FastGPT version and the deployed MongoDB version.
- Initiate retrieval for EL detection images and IV data, and confirm that the returned results include corresponding visual content and numerical indicators.
The values provided are common starting points and should be measured against the reader's own samples.