What the data for this category looks like
The main entities in batch fine chemical synthesis are enterprises producing coatings, resins, additives, agrochemical intermediates, and specialty chemicals. Data sources include formula and process libraries, raw material batch records, reactor DCS data, weighing system logs, LIMS test reports, and warehouse inventory information. Daily documents include batch production instructions, feeding verification sheets, reaction process records, cleaning records, test reports, deviation and change orders. Data is updated per production batch or each critical step. Most documents are structured forms and unstructured process records, with fields including batch number, feeding parameters, reaction conditions, test indicators, and more. Some data includes units such as kilograms, degrees Celsius, hours, and there is no unified universal format.
What constraints these characteristics impose on the knowledge base retrieval and recall link
The multi-batch, multi-step update nature of this category leads to data fragmentation, requiring retrieval to associate and match relevant documents by batch. The mixed structured and unstructured characteristics of different documents require support for both semantic retrieval and precise field matching. Data updated per batch requires filtering invalid batch ranges during retrieval to avoid interference from cross-batch documents. The need for formula confidentiality requires that retrieval results strictly limit access permissions and must not disclose core process details. Documents from critical audit links such as feeding, reaction, and cleaning are highly correlated, requiring recall of associated content in process order to improve audit efficiency.
The following configurations are adapted for FastGPT v4.17.0.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the semantic accuracy of fine chemical process documents, filtering low-relevance general chemical documents |
limit | Top 6 entries | Controls the number of documents recalled per single retrieval, adapting to the total volume of standard documents for single-batch audits |
searchMode | embedding | Performs semantic matching for unstructured reaction records and cleaning records, while supporting structured field retrieval |
usingReRank | Enabled | Reranks recalled results to prioritize documents from critical audit links such as feeding and reaction |
datasetSearchUsingExtensionQuery | Enabled | Expands retrieval based on audit keywords used by users, for example, expanding "feeding verification" to "feeding verification sheet" and "batch feeding record" |
collectionIds | Bind corresponding datasets by audit scenario | Differentiates supplier qualification documents and internal production process documents to avoid interference from cross-scenario retrieval |
Three common mistakes
- Phenomenon: After selecting the corresponding knowledge base, there are no optional values in the reference variable dropdown menu. Cause: Fields of structured documents have not been mapped to retrievable reference variables, or the field indexing function of the dataset has not been enabled.
- Phenomenon: After retrieval results are returned, the front end cannot display them in document name order, and there is no content entry when clicked. Cause: Metadata return fields for retrieval results have not been configured, or the front-end document detail interface has not been connected.
- Phenomenon: The number of retrieval results does not match the configured
limitparameter, or irrelevant basic chemical documents appear. Cause: Thesimilaritythreshold has not been set correctly, or the retrieval dataset range has not been limited viacollectionIds.
How to confirm the configuration is correct
- Upload a batch of test feeding verification sheets and reaction process records, execute a retrieval with the keyword "202405 batch feeding verification", and check whether the document types of returned results match the keyword.
- Adjust the
similarityparameter range to verify whether the relevance change of returned results meets expectations. - Check that retrieval results only include documents from datasets corresponding to the currently bound
collectionIds, with no cross-dataset interference. - After enabling
usingReRank, check whether reranked results prioritize documents from critical audit links such as feeding and reaction.
The values provided are common starting points and should be measured against the reader's own samples.