What This Category of Data Looks Like
Data sources for this category include formula and process libraries, raw material batch records, reactor DCS data, weighing system logs, LIMS systems, and warehouse inventory data. Daily documents include batch production instructions, feeding review sheets, reaction process records, cleaning validation records, inspection reports, and deviation and change orders. Each document includes fields such as batch number, feeding amount, reaction temperature, and residence time. Units include kilograms, degrees Celsius, hours, and other standard units. Data is updated once per production batch or key process step. Due to multi-product production switching, a single batch of data is linked to multiple sets of process and cleaning records. Accurate matching of context information for production nodes is required.
Constraints on Knowledge Base Retrieval and Recall
Multi-batch, multi-dimensionally associated process data requires retrieval to support combined filtering by multiple conditions such as batch number, process step, and equipment number. This prevents irrelevant cross-batch data from interfering with recall results. Dynamic data updated per single batch or key step requires retrieval to match context for real-time process nodes. This supports immediate query needs during production. Documents contain confidential formulas and sensitive process parameters. Retrieval results must pass permission checks, and only content within authorized scope is returned. Multi-product production switching creates cleaning validation requirements. Retrieval must associate cleaning records for different batches on the same equipment. It must also preserve the original hierarchical logic of documents to avoid structural loss.
Configuration Setup
This document was written using FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Fine chemical process parameters have high precision requirements. Low-correlation general process documents must be filtered out, and only results that meet matching thresholds and include batch information are retained. |
limit | Top 6 entries | The number of documents linked to a single batch process is limited. Excessive recall will exceed the context window and reduce the accuracy of large model analysis. |
searchMode | mixedRecall | Semantic matching of process parameters and keyword matching of batch numbers, equipment numbers, and other fields are required. This supports multi-condition retrieval needs. |
usingReRank | Enabled | Process document similarity can be disrupted by batch numbers. Reranking corrects sorting biases from initial recall, and prioritizes results associated with the current batch. |
datasetSearchUsingExtensionQuery | Enabled | Production problem descriptions often include vague statements. Extended queries can supplement keywords such as associated process steps and inspection indicators, improving recall coverage. |
collectionIds | Knowledge base collection IDs bound to the current production batch | Retrieval scope must be limited to the process, inspection, and cleaning document collections linked to the current batch. This avoids interference from irrelevant cross-category data. |
Three Common Misconfigurations
- Symptom: Retrieval results lose the hierarchical logic of document titles and subordinate categories. Corresponding process steps cannot be associated. Cause: The title hierarchy of markdown documents is not preserved. Data is only split by paragraphs, leading to broken semantic associations.
- Symptom: Extra spaces appear between numbers and text in retrieval results, making process parameters harder to read. Cause: Format residues from original documents are not addressed during knowledge base splitting. Display rules for numbers and units are not aligned.
- Symptom: Real-time process data from reactor DCS cannot be recalled. Cause:
collectionIdsis not configured to limit the retrieval scope to the bound real-time data collection, leading to retrieval covering irrelevant documents.
How to Verify Correct Configuration
- Enter a specified batch number and process step keyword, then check if retrieval results only return documents associated with that batch.
- Upload a markdown process document with title hierarchy, then verify that retrieval results retain the subordinate relationship between titles and content.
- Test vague problem descriptions such as "a batch reaction anomaly", then check if associated deviation and change order documents can be recalled.
- View the similarity score of retrieval results, confirm that the score falls within the preset
similarityrange.
The values provided are common starting points and should be evaluated against the reader's own samples.