What data for this category looks like
Data sources for this category include ERP material master data, melting furnaces and thermometers, spectrometers, MES heat records, mold management systems, and customer drawing technical documents. Daily documents cover foundry process cards, batching lists, heat records, mold histories, first-piece inspection records, material test reports, and nonconforming product review forms. Data updates occur per heat, per batch, or daily. Fields include heat number, melting temperature, pouring duration, mold usage count, defect type code, and material test values. Units include degrees Celsius, minutes, counts, and similar units.
What constraints these characteristics impose on knowledge base retrieval and recall
This category has high data update frequency, with associated fields tied to individual heats and batches. Knowledge base retrieval and recall must support precise matching via fields such as heat number and mold number, to avoid retrieving historical data from unrelated batches. The coexistence of multiple document types requires retrieval to adapt to structured data such as batching lists and unstructured data such as process cards, ensuring content from all sources can be effectively retrieved. Data volume grows dynamically with production pace. The number of single-recall results must be limited to prevent exceeding the model's context processing limit, while supporting incremental sync updates for heat records and mold history data. Business workflows require precise verification of alloy batching and return charge ratios, and tracing defects to their corresponding heat and mold numbers. Retrieval results must have strong contextual association capabilities, rather than returning only isolated document fragments.
How to set configurations
Choose a retrieval mode through representative-query testing; configure category, batch, or entity filters separately because the mode itself does not enforce those boundaries.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category requires precise matching of core business fields such as heat numbers and batching ratios. A threshold that is too low will introduce data from unrelated batches, while a threshold that is too high will fail to retrieve maintenance solutions for similar defects. |
limit | Top 8–12 results | Documents related to foundry work have long individual content. Too many recall results will exceed context limits, while too few will fail to cover complete maintenance and spot check workflows. |
usingReRank | Enabled | This category has high relevance requirements for data. Reranking can filter low-relevance recall results, improving matching accuracy between maintenance solutions and spot check questions. |
datasetSearchUsingExtensionQuery | Enabled | User spot check questions such as "How to handle shrinkage porosity in a specific heat" can be expanded into query terms that include heat numbers and material parameters, improving retrieval accuracy. |
collectionIds | Bind to the dedicated foundry equipment spot check and maintenance dataset | Isolate knowledge base data from other categories, avoiding retrieval of unrelated forging or machining documents, and ensuring the retrieval scope aligns with business scenarios. |
Note: The configurations on this page are based on FastGPT v4.17.0
Three common misconfigurations
- Symptom: Empty results or permission errors when calling the API to reference a private knowledge base. Cause: The
collectionIdsparameter is not configured correctly, and the target dataset is not added to the allowed call collection list. - Symptom: Crashes on the chat page or knowledge base page. Cause: The
limitparameter value is not restricted, and too many documents are recalled in a single request, causing context overflow and triggering platform memory limits. - Symptom: Corresponding heat batching lists cannot be retrieved after creating a new data index. Cause:
datasetSearchUsingExtensionQueryis not enabled, so the heat number in the user's question cannot be automatically associated with index fields, leading to retrieval matching failure.
How to confirm configurations are set correctly
- Run a retrieval for a single-heat spot check question, verify that the recall results only include data such as batching and melting parameters for the corresponding heat.
- Check the number of results returned by the retrieval, confirm that it falls within the range set by the
limitparameter. - Call the API to test referencing the private knowledge base, confirm that the returned results include the expected maintenance solutions and spot check documents.
- Check that the reranking function is active, confirm that highly relevant documents appear at the top of the retrieval results.
The configuration values provided are common starting points, and should be evaluated against the reader's own samples.