What Data for This Category Looks Like
The main entities in this category are color printing, label, carton, flexible packaging, and pharmaceutical packaging processing enterprises. Daily generated documents include pre-press files, imposition sheets, printing process sheets, ink color matching records, first article signing samples, barcode inspection, die-cutting and lamination records, etc. Data sources include customer manuscripts, color management systems, printing press operation data, ink and substrate batch records, spectrometer and barcode detector outputs. Data updates are triggered per order, batch, or plate change. Each individual document contains three types of fields: text descriptions, numerical parameters, and inspection results. There are no unified fixed units, and matching must be done according to specific business scenarios.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
The data characteristics of this category impose multiple constraints on the retrieval and recall link. Frequently updated business data requires the knowledge base to support incremental synchronization and version management, to avoid recalling expired process sheets or inspection records. Multiple types of documents coexist, so both text semantic matching and precise numerical parameter matching must be supported, to avoid missing key registration and barcode requirements. The scale of documents involved in a single task is limited, but accuracy requirements are extremely high. Too many recall results increase the review burden for subsequent drafting, while too few may miss key process requirements. Incorrect recall will directly lead to serious consequences such as misprint scrap, so the relevance threshold for recall must be strictly controlled.
How to Configure
This page’s configuration is based on FastGPT v4.17.0. The following are specific configuration items and their basis:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
Similarity Threshold (similarity) | 0.75–0.82 | Printing process parameters and inspection results require high matching accuracy, to avoid recalling irrelevant ink color matching records or imposition sheets |
Single Search Limit (limit) | Top 8 entries | Process documents for this category have compact content. Too many recall results increase the context processing burden and affect drafting efficiency |
Query Optimization (datasetSearchUsingExtensionQuery) | Enabled | Questions for process drafting and exception reports often use simplified expressions. Optimization can match the core content of complete process sheets, first article signing samples, and other documents |
Re-Ranking (usingReRank) | Enabled, take top 12 candidates for re-ranking | Recalled documents must be sorted by relevance, to ensure key content such as first article signing sample requirements and registration rules is returned first |
Retrieval Mode (searchMode) | Hybrid Retrieval | Supports both text semantic matching and numerical parameter matching, adapting to the document structure of this category which includes both text descriptions and inspection data |
Collection ID Filter (collectionIds) | Configured by document type | Distinguishes different business collections such as pre-press files, inspection records, and batch archives, to avoid recalling irrelevant documents across types |
Three Common Mistakes
- Phenomenon: After enabling
usingReRankanddatasetSearchUsingExtensionQuery, API call response time exceeds 30 seconds. Cause: The number of single recall entries and the scale of re-ranking candidate sets are not limited. A frequently updated knowledge base leads to overload of retrieval calculation volume. - Phenomenon: Recall results do not include documents with the specified file name. Cause:
collectionIdsfor the corresponding business collection is not configured, or the file name metadata field is not associated in the retrieval parameters. - Phenomenon: When a locally deployed chatbot calls the API, the returned content does not include knowledge base fragments. Cause: The knowledge base retrieval switch is not enabled, or the API request does not carry the correct
collectionIdsparameter.
How to Confirm the Configuration Is Correct
- Go to the knowledge base configuration page in the FastGPT backend, check whether the switch status of
usingReRankanddatasetSearchUsingExtensionQuerymatches the preset configuration. - Initiate a test query, enter "First article signing sample requirements", and verify whether the returned results include the key content of the first article signing sample document.
- Check the API response log, confirm that the returned context fields include the recalled document fragments, and that the response time meets the preset threshold.
- Upload a new die-cutting record document, initiate a query containing "die-cutting accuracy", and confirm that the new document is correctly recalled.
The values given are common starting points and should be measured against the reader's own samples.