This document is written based on the configuration logic of FastGPT v4.17.0.
What the data for this category looks like
This category’s data typically comes from silicon wafer incoming material records, production MES systems, real-time operating data from diffusion furnaces, coating equipment, and printing presses, EL and IV test results, and laboratory quality analysis documents. The update rhythm follows: each batch of materials corresponds to one process ledger, detection data for each cell is uploaded in real time, and daily summaries of daily yield and process reports are compiled. Daily documents include process cards, diffusion and coating parameter tables, printing and sintering process records, EL defect detection images, IV performance grading data, and yield analysis reports. Fields cover process temperature, coating thickness, defect coordinates, conversion efficiency, and more. Units include degrees Celsius, micrometers, and others.
Constraints imposed on knowledge base retrieval and recall
Photovoltaic cell manufacturing ledger data has high update frequency and fine granularity. Real-time upload of detection data for each cell leads to frequent incremental knowledge base updates. If retrieval configurations are not adapted to high-frequency updates, recalled data may lag. Data includes two categories: structured parameters and unstructured images. A single retrieval mode cannot cover all document types, so both matching logics must be supported. Wafer-level data has strong relevance. A single batch’s ledger will link to multiple sets of equipment parameters and test records. Retrieval must support cross-document associated recall, otherwise key process information may be missed.
Configuration settings
Use the following table for recommended configurations:
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Photovoltaic process parameters and detection data have high matching accuracy requirements. This interval filters invalid recall results with low matching degrees |
limit | Top 10 entries | A single batch of production ledgers usually has no more than 10 associated documents. This value balances recall completeness and context load |
usingReRank | Enabled | Semantic similarity of EL detection images and IV performance data needs secondary calibration to improve ranking priority for highly matched documents |
datasetSearchUsingExtensionQuery | Enabled | Production ledgers often use batch numbers to link multiple document types. Extended queries cover all data associated with a batch |
searchMode | mixedRecall | Balances keyword matching for structured parameters and vector matching for unstructured image descriptions to adapt to multiple document types |
Three common configuration mistakes
- Phenomenon: Retrieval response times out after knowledge base content grows. Cause: The
limitparameter is not adjusted, leading to too many low-priority recalled documents and increased context loading time. - Phenomenon: Recalled results do not display the corresponding document file names. Cause: Document metadata index configuration is not enabled, or only pure vector retrieval is used without binding file name association logic.
- Phenomenon: API call responses do not include knowledge base-associated ledger data. Cause: The target knowledge base collection is not bound in the chat application configuration, or the
collectionIdsparameter is not specified in the API request.
How to verify correct configuration
- Upload a single photovoltaic process card and EL detection image, initiate a query that includes process parameters and defect descriptions, and confirm recalled results include the target document’s file name and corresponding content
- Simulate high-frequency data import, send batch retrieval requests, and check that response times meet measured thresholds
- Call the API interface with the correct
collectionIdsparameter, and verify returned results include production ledger data associated with the knowledge base - After enabling
usingReRank, compare recall result rankings before and after enabling, and confirm highly matched process parameter documents appear higher in the list
The values provided are common starting points and should be measured against the reader's own samples.