What Data for This Product Category Looks Like
Daily documents for complete household appliance manufacturing include configuration BOMs, assembly work instructions, refrigerant charging records, electrical safety test reports, performance spot check reports, and rework records. Data sources cover ERP/MES systems, component batch ledgers, tightening and charging equipment logs, safety compliance tester data, performance laboratory records, and after-sales fault feedback data. Data updates follow a pattern of single records per product, shift-by-shift data aggregation, and batch report generation per spot check plans. Fields include model numbers, batch numbers, process parameters, test results, fault descriptions, and more. Field formats vary across documents from different sources.
Constraints on Vector Models and Indexing
Mixed-line assembly of multiple product models requires vector indexes to distinguish assembly processes and parameter sets for different models, preventing vector confusion across model types. High-frequency updates of per-product and per-shift data require indexes to support incremental updates, reducing resource consumption from full reindexing. Non-standardized fields from multiple sources require retaining key identifier fields during chunking, ensuring vector matching can link to production data for the corresponding batch. The high-risk consequences of safety compliance and functional anomalies require the confidence threshold of vector matching to align with business needs, avoiding false recall of low-relevance historical data.
How to Configure Settings
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Aligns with the length of single-process descriptions and test records for complete household appliance documents, avoiding splitting critical associated information |
| Overlap Length | 100–150 characters | Connects key fields such as batch numbers and process parameters across adjacent chunks, ensuring information association across chunk boundaries |
| Similarity Threshold | 0.75–0.85 | Distinguishes subtle process differences between batches of the same model, preventing false recall of defect records from unrelated batches |
maxToken | 8192 | Aligns with the context limit of mainstream embedding models supported by FastGPT v4.17.0, covering complete assembly process records |
defaultToken | Set from representative sample documents | defaultToken is the default token count for direct text chunking; tune it on representative source documents. See the v4.17.2 model-field reference. It is not a vector dimension, embedding-model input cap, or retrieval/LLM context. |
| Maximum Search Results Per Query | Top 10 entries | Covers multi-batch associated retrieval needs in mixed-line production, avoiding interference from redundant data during defect analysis |
Three Common Configuration Errors
- A 503 error appears with the prompt "No available channel for model text-embedding-v3 under current group default". The cause is failing to separately specify an embedding model for the indexing process, and reusing the default API channel which causes a call failure.
- Changes to the CHATAPIKEY environment variable do not take effect. The symptom is a key verification failure prompt during retrieval. The cause is failing to rebuild the docker-compose container, so the new environment variable is not loaded.
- Retrieval result matching accuracy is low. The symptom is recalled defect records that cannot be linked to assembly data for the corresponding batch. The cause is failing to retain key fields such as component batch numbers and tightening torque during chunking, resulting in missing vector information.
How to Confirm Successful Configuration
- The index configuration page of the FastGPT knowledge base is accessed, and confirmation is made that the embedding model has been separately specified as the target model, with the default aggregation channel not used.
- A test retrieval is performed, and it is checked that the number of returned results matches the set maximum search results per query, with no abnormal truncation.
- The vector index update logs are checked, and confirmation is made that newly uploaded assembly records and rework records have been correctly indexed, with no field loss prompts.
- The docker-compose container is restarted, and verification is made that the CHATAPIKEY configuration takes effect, with no key verification failure prompts during retrieval.
The values provided here are common starting points and should be evaluated against the reader's own samples.