What Data Looks Like for This Category
Daily documents for integrated iron and steel enterprises cover iron smelting, steelmaking, continuous casting, quality control, and scheduling departments. These documents include heat plans, batching lists, converter smelting records, composition reports, continuous casting process records, and slab judgment sheets. Data sources include hot metal and scrap steel data, L2 process control systems, laboratories, MES, energy medium systems, and equipment status. Documents are generated per heat, synchronized in real time, and summarized and archived per shift. Documents contain fields such as composition values, process parameters, and equipment status, with structure adjusted dynamically based on operation links.
Constraints on Vector Models and Indexes
High-frequency, small-batch data generated in real time per heat requires indexes to support low-latency incremental updates without disrupting operation rhythms. Mixed input of multiple structured document types requires vector models to adapt to semantic encoding for different fields, ensuring consistent association of composition, process, and equipment data. Strong coupling between heats and casting runs requires indexes to establish cross-document associated indexes, enabling quick location of associated operation data during cross-document reviews. Scattered data sources with frequent updates require a unified index entry and permission control to prevent data omission or duplicate indexing.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the average length of single heat documents, avoids semantic breaks across fields |
| Overlap Length | 100–150 characters | Retains associated semantics of adjacent operation links, prevents chunking from cutting continuous casting process context |
| Similarity Threshold | 0.75–0.85 | Filters low-relevance historical documents, focuses on core review data for the current heat |
| Single Search Limit | Top 10 results | Matches the upper limit of document count for single-heat cross-document reviews, avoids interference from redundant information |
| Reranking | Enabled | Sorts multi-source heterogeneous documents, prioritizes matching composition and temperature parameters relevant to reviews |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
defaultToken | 512 | Serves as the default token base for single-document chunks, adapts to encoding needs for small-field documents |
Three Common Misconfigurations
- Phenomenon: Online recall test shows no valid results. Reason: Failed to bind the reranking model synchronously during the indexing stage, or the input length of the reranking model does not match the current document chunk length.
- Phenomenon: Field empty error occurs during index construction. Reason: Failed to filter temporary cache fields in L2 process control data, leading to empty values that cannot generate valid vectors.
- Phenomenon: Recall results contain a large number of cross-heat redundant documents. Reason: Failed to limit the search scope to associated data of the current operation heat, and the similarity threshold is set too low.
How to Confirm Proper Configuration
- Upload a single heat composition report and continuous casting process record, run an online recall test, and check that the returned results include the corresponding documents and the sorting meets expectations.
- View index construction logs, confirm that there are no error messages about field empty values or exceeded chunk length limits.
- Adjust the similarity threshold, verify that the number of recall results changes as expected in line with the threshold adjustment.
- After configuring
maxTokento 4096, merge three or more associated documents, confirm that the context loads without truncation.
The values provided are common starting points and should be measured against the reader's own samples.