What Data Looks Like for This Category
Typical entities in this category are manufacturers of coatings, resins, additives, agrochemical intermediates, and specialty chemicals. Daily generated documents include batch production instructions, material dosing review slips, reaction process records, cleaning records, inspection reports, deviation and change orders. Data sources cover formula and process libraries, raw material batch information, reactor DCS data, weighing systems, LIMS systems, and warehouse systems. Data is updated per production batch or each key process step. Due to multi-variety production with batch switching, documents and data from different batches must be strictly distinguished. Permission controls related to formula confidentiality are also part of the requirements.
Constraints Imposed on Knowledge Base Retrieval and Recall
Multi-batch and multi-document type features require retrieval to accurately match by batch and process step, to avoid mixing across batches. Frequently updated step-by-step data requires the retrieval pipeline to support incremental synchronization. This ensures recalled content matches the latest data of the current production link. Mixed structured and unstructured data from multiple sources requires retrieval to support documents and system exported data in different formats. It also requires support for filtering by field conditions such as raw material batch number and reactor number. The scenario of multi-variety production switching requires retrieval to support quick switching of knowledge base collections. This adapts to the document scope of the current production category, while meeting permission requirements for formula confidentiality.
Configuration Settings
This document is written based on FastGPT v4.17.0, and the following configurations are adjusted for the retrieval module of this version.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the rigor requirements for matching fine chemical process documents, filters low-relevance general descriptions, and retains content strongly related to specific dosing and reaction steps |
limit | Top 8 entries | Covers the number of core document types for a single production batch, avoiding context overload that affects large model processing |
searchMode | embedding | Adapts to multi-source mixed process documents, prioritizes recalling semantically matched content, and supports unified retrieval of structured and unstructured data |
usingReRank | Enabled | Re-ranks initial recalled results, corrects deviations from single similarity calculations, and ensures high-priority documents such as reaction process records and material dosing review slips appear at the top |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands query terms based on technical review questions, supplements retrieval for key fields such as process parameters and batch numbers, and improves retrieval accuracy |
collectionIds | Collection IDs bound to the current production category | Only retrieves knowledge base content corresponding to the current production batch, avoids interference from cross-category documents, and meets formula confidentiality requirements |
Three Common Misconfigurations
- Symptom: An error occurs when calling after adding a knowledge base, and the log shows field parsing failure or context overrun. Cause: The number of recalled entries and re-ranking parameters are not adjusted for multi-type, multi-batch fine chemical documents, causing the total length of recalled documents to exceed the large model's supported range.
- Symptom: Retrieval results include documents from non-current production batches, with batch number fields empty or mismatched. Cause: The
collectionIdsparameter is not configured correctly, the retrieval scope covers irrelevant knowledge base collections, and document content is not filtered by production batch. - Symptom: Responses do not strictly reference process content from the knowledge base, and descriptions outside the knowledge base appear. Cause: A reasonable
similaritythreshold is not set, orusingReRankis not enabled, resulting in recalled low-relevance general documents that interfere with large model output.
How to Confirm Configuration Is Correct
- Manually enter a material dosing review question for a single batch, check if recalled document types include core documents such as material dosing review slips and batch production instructions.
- View the retrieval log to confirm that only knowledge base content corresponding to the currently bound
collectionIdsis returned, with no cross-category or cross-batch documents. - Adjust the value range of
similarity, verify whether the relevance of retrieval results meets the rigor requirements for process review. - After enabling
usingReRank, check whether high-priority documents such as reaction process records and cleaning records appear at the top of retrieval results.
The values provided here are common starting points and should be evaluated against local test samples.