What data for this category looks like
The data for this category comes from engineering orders, sand, cement and admixture inspection records, aggregate moisture content test reports, mixing station control system logs, and project laboratory archives. Update frequency changes with the production process: Each truckload of concrete generates a corresponding production and transport order, each batch of raw materials entering the site generates raw material inspection records, and test block records and strength reports are summarized daily. Documents include mix proportion notices, opening identification forms, test block strength reports, etc. Fields cover batching parameters, inspection values, production time, transport scheduling information, with slight differences based on mixing station and laboratory record habits.
What constraints do these characteristics impose on the "model access and configuration" link
The frequently updated production and inspection data for this category requires the model access link to support incremental index configuration, avoiding excessive resource usage from full reindexing. Dispersed document fields and non-uniform formats require limiting the range of recalled document types during configuration, reducing interference from irrelevant content. Single production data volume is limited, but overall data scale grows with orders, requiring adaptation of a reasonable context carrying limit. At the same time, process and standard Q&A involve quality responsibility, so configuration must ensure the accuracy of recalled content, avoiding missing or incorrect recalled content due to improper parameters.
How to set the configuration
The following configurations are formulated based on FastGPT v4.17.0.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Adapts to the single-batch data volume and multi-document splicing requirements of ready-mixed concrete production documents, avoiding context overflow |
quoteMaxToken | 1500–2000 token | Limits the reference length of a single recalled document, matching the core content length of single documents such as mix proportion sheets and test block reports |
temperature | 0.1–0.3 | Ensures the accuracy of answers for process and standard Q&A, avoiding generated content that deviates from actual production requirements |
AIPROXY_API_ENDPOINT | Fill in the local Ollama API address according to the deployment environment | Adapts to local deployment model call requirements, ensuring correct model access path |
AIPROXY_API_TOKEN | Fill in the corresponding key according to deployment configuration | Completes identity verification for local models, ensuring legitimate call permissions |
maxTokens | 2000–3000 token | Limits the output length of a single round of answers, avoiding overly long process descriptions that exceed the reading requirements of engineering application scenarios |
Three common configuration mistakes
- Rerank model is not triggered during online recall testing. The cause is that the Rerank configuration enabled in the knowledge base index link is not associated with the current application's Q&A process, or the deployment path of the Rerank model is filled incorrectly.
- Context is not sent to the large model after
maxContextexceeds the threshold. The cause is that the application's context carrying limit does not match themaxContextparameter, or the system has an implicit limit on the total length of a single round of context, and content beyond the limit is automatically truncated without being passed. - Request Time errors occur frequently during conversations. The cause is that the response timeout setting of the local Ollama model is too short, or the network delay of
AIPROXY_API_ENDPOINTis too high, and a reasonable timeout retry mechanism is not configured.
How to confirm the configuration is complete
- Enter the recall test interface of the knowledge base, enter process-related questions, and check whether the number and content of recalled results match the preset document range.
- Adjust the
temperatureparameter and initiate a test Q&A, observe the change in the rigor of the answer, and confirm that the parameter takes effect. - Check the model call log to confirm that there are no authentication or path errors in the call requests for
AIPROXY_API_ENDPOINTandAIPROXY_API_TOKEN. - Configure
maxContextto the upper limit value and initiate a long-context Q&A, confirm that the large model can fully receive the spliced recalled content without truncation prompts.
The values provided are common starting points and should be measured against the reader's own samples.