What this category of data looks like
Data for this category comes from formula systems, internal mixers, vulcanizers, and laboratory testing equipment of rubber product enterprises such as tire, seal, and hose manufacturers. Daily documents include formula sheets, weighing records, mixing curves, vulcanization process cards, finished product performance reports, and more. Each document is bound with unique identifier fields tied to raw material batches and production mold runs. Data updates sync with each batch of raw materials, each batch of rubber compound, and each production mold run. Fields include rubber compound ratio, additive dosage, mixing temperature range, vulcanization time and pressure parameters, with units mostly degrees Celsius, minutes, and megapascals.
What constraints these characteristics impose on knowledge base retrieval and recall
High-frequency incremental updates require the knowledge base to support fast synchronization of new documents by batch and mold run, to avoid resource consumption from full index rebuilding. Multi-type document structures with unique identifiers require retrieval to support exact matching of raw material batches and production mold runs, while also supporting range queries for numerical parameters. Most fields in single documents are process parameters, so highly relevant entries matching the current process scenario must be prioritized for recall, to avoid irrelevant documents consuming context quota. Document formats and structures vary across different equipment, so retrieval configuration must support mixed recall logic for structured parameter tables and time-series mixing curves.
The following configuration items are formulated based on FastGPT v4.17.0.
How to set configurations
| Configuration Item | Suggested Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets precision requirements for matching process parameter text, to avoid recalling irrelevant mixing or vulcanization records |
limit | Top 6 entries | Adapts to context quota limits for single-round chats, while covering associated reference documents for typical process scenarios |
searchMode | mixedRecall | Adapts to mixed document types including structured parameter tables such as vulcanization process cards and time-series curves such as mixing curves |
usingReRank | Enabled | Performs secondary sorting on initially recalled entries by process relevance, to optimize recall priority of key parameters |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonyms for process questions, for example associating "under-cure" with documents related to "insufficient vulcanization" |
collectionIds | Collection IDs grouped by business domain | Isolates different knowledge base collections such as mixing, vulcanization, and finished product testing, to avoid cross-domain recall interference |
Three Common Misconfigurations
- Phenomenon: When calling knowledge base chat, the
<Reference>module in the system prompt does not display any retrieved document content. Cause: The "inject recalled content into system prompt" switch was not enabled in chat configuration, causing retrieval results to not be correctly injected into the context. - Phenomenon: Unable to export the filename list of uploaded files in the knowledge base. Cause: The knowledge base metadata collection switch was not enabled, causing file information such as filenames to not be synchronized to the index library, so they cannot be retrieved or exported.
- Phenomenon: Recalled documents include process records from other non-target categories. Cause: The
collectionIdsparameter was not configured, causing knowledge base collections across business domains to be retrieved simultaneously, introducing irrelevant content.
How to Verify Correct Configuration
- Upload a test vulcanization process card, submit a question including "the vulcanization parameters of this mold", confirm whether recall results include fragmentary content of the document.
- View the system prompt area in the chat interface, confirm that the
<Reference>module displays matched document filenames and specific parameter fragments. - Adjust the
similarityparameter to 0.7, submit a question including mixing temperature requirements, check changes in the relevance of recall results to verify parameter effectiveness. - Export the metadata list of the knowledge base, confirm that the uploaded test document filename has been correctly included and can be retrieved.
The values provided are common starting points and should be measured against samples for the intended use case.