This document is written based on FastGPT v4.17.0.
What the Data for This Category Looks Like
Data for this category comes from customer drawings, special characteristic lists, ERP/MES systems, measuring tools, equipment process parameters, and supplier quality certificates. Daily documents include control plans, process flow diagrams, work instructions, first article and patrol inspection records, SPC reports, PPAP materials, and 8D reports. Data is continuously collected and updated by production batch and shift. Some documents are linked to specific process numbers, equipment numbers, and batch numbers. Fields include special characteristic parameters, process capability indicators, abnormal disposal records, and more. The word count of individual documents varies widely.
What Constraints Do These Characteristics Impose on Knowledge Base Retrieval and Recall?
Data for this category is multi-source and has high real-time requirements. It must support real-time data access from multiple types of systems. Document types are diverse, with significant structural differences. Some documents contain both structured parameters and unstructured descriptions. Retrieval must balance semantic matching and precise field matching. Individual documents have wide word counts and are linked to production batch information. Retrieval must support precise filtering by process, equipment, batch, and other dimensions. It must also adapt to chunking and recall logic for different documents. The high-frequency update requirement means the retrieval pipeline must support incremental synchronization. This avoids resource consumption from full reindexing.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Automotive parts quality documents mostly contain precise parameters. Low-relevance results must be filtered to avoid mixing in abnormal records from non-target processes |
limit | Top 8 entries | Mass production scenarios require covering associated records across multiple processes and batches. Too many results will exceed the context window, while too few will miss critical root cause information |
searchMode | mixedRecall | Documents for this category contain both structured parameters and unstructured descriptions. Hybrid retrieval balances semantic matching and precise field matching |
usingReRank | Enabled | Abnormal analysis for mass production requires precise matching of 8D reports and SPC reports linked to root causes. Reranking improves the ranking priority of highly relevant results |
datasetSearchUsingExtensionQuery | Enabled | This category requires linking fields such as special characteristics and process numbers. Extended queries can supplement retrieval of relevant upstream and downstream documents |
collectionIds | Bind corresponding datasets by production batch | Data for this category is updated by batch. Binding datasets enables precise recall of production documents for the corresponding batch, avoiding interference across batches |
Three Common Misconfigurations
- Phenomenon: Retrieval results do not include core paragraphs of long documents, or return empty results. Cause: The chunking logic for long documents is not adapted. If chunking is done directly with a fixed length, key segments of long documents are discarded because they exceed the length limit supported by the vector model, leading to recall failure.
- Phenomenon: The latest batch records cannot be synchronized after the knowledge base is updated, requiring manual re-upload of all documents. Cause: Incremental update trigger rules are not configured. Only full upload is supported, so partial content updates for individual documents are not possible, which cannot adapt to the high-frequency batch-level update requirements of this category.
- Phenomenon: AI conversation output delay exceeds the preset threshold. Cause: The
usingReRankreranking function is not enabled, or thelimitvalue is set too high, leading to too many recalled results. Exceeding the context window requires additional processing, which slows down output speed.
How to Verify Proper Configuration
- Upload a typical 8D report and SPC report, perform an abnormal retrieval, and check whether the results include key paragraphs of the corresponding documents.
- Simulate a single abnormal retrieval, check whether the number of returned results matches the preset
limitvalue range. - Submit a locally modified work instruction, verify whether the knowledge base automatically synchronizes the updated content without requiring a full re-upload.
- Check the overall retrieval and conversation time, confirm that the output speed meets the response requirements of the business scenario. Adjustments can be made by modifying the
similaritythreshold orlimitvalue.
The values given are common starting points and should be measured against the reader's own samples.