What this category of data looks like
Typical organizations in this category are manufacturers of pumps, valves, compressors, reducers, machine tools and non-standard equipment. Daily generated documents include assembly process cards, BOMs, torque records, gap adjustment records, commissioning reports, factory inspection sheets and packing lists. Data sources cover PLM/ERP systems, part inspection data, assembly tool collection terminals, test benches, MES systems and customer technical agreements. Documents are updated once per piece of equipment or per order. Document fields include precise numerical values such as torque value (unit: Nm), gap value (unit: mm), coaxiality tolerance (unit: μm), as well as text content such as assembly steps and material codes.
What constraints these characteristics impose on knowledge base retrieval and recall
Data sources for this category are scattered and updated frequently. This requires knowledge base retrieval to support associated recall of multi-system data, while ensuring real-time data synchronization. Document structures vary widely: there are hierarchical BOM documents, as well as inspection data consisting of single numerical records. Retrieval must match different types of content features. The matching accuracy of precise numerical fields directly affects assembly quality. The relevance of retrieval results must meet the strict requirements of assembly audits. In scenarios with parallel assembly of multiple orders, irrelevant documents across orders or equipment must be avoided from being recalled. Otherwise, audit personnel may make misjudgments.
How to configure settings
This configuration applies to FastGPT v4.17.0. Specific parameter configurations are as follows:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches precise numerical documents such as torque and gap values, filters low-relevance general process content |
limit | Top 6 entries | Covers multiple types of documents required for single-unit assembly such as process documents, records and BOMs, while controlling the total context length |
usingReRank | Enabled | Performs secondary ranking on recalled inspection records and process cards, prioritizes key content such as coaxiality and preload force |
datasetSearchUsingExtensionQuery | Enabled | Supplements retrieval of associated fields such as order numbers and equipment models, matches query requirements for BOM and physical kit alignment |
collectionIds | Configured by assembly stage | Distinguishes between process libraries, inspection record libraries and BOM libraries, prevents irrelevant content across stages from being retrieved |
Three common misconfigurations
- Scenario: A query targeting only the gap adjustment record of a single reducer recalls assembly records of other equipment under the same order. Cause:
collectionIdswas not used to limit the dataset for the current assembly batch, resulting in retrieval of irrelevant documents across equipment. - Scenario: FastGPT deployed online displays garbled text for imported assembly process cards, while local deployment has no such issue. Cause: The default file encoding of the online server does not match the GBK or UTF-8 encoding of the documents, and no encoding format was specified during the import process.
- Scenario: The latest torque record data cannot be found in retrieval results. Cause: No incremental synchronization mechanism was configured, so the knowledge base did not update the latest assembly tool collection data in a timely manner.
How to confirm the configuration is correct
- Enter a BOM hierarchy query for a single device, verify that recalled results only include BOM documents for that device, and adjust the
similarityandlimitparameters to a range that meets business requirements. - Import a batch of assembly process cards and torque records, confirm that imported document content online matches the local version, and verify that the encoding format configuration is correct.
- Simulate a query for an inspection audit scenario, check that recalled results rank gap adjustment and coaxiality related documents at the top, and confirm that the
usingReRankconfiguration is enabled. - Enter a query for a specific order number, verify that recalled results only include assembly records under that order, and confirm that the
collectionIdsgrouping configuration is correct.
The values provided are common starting points and should be measured against the reader's own samples.