What the data for this category looks like
Daily documents for general mechanical equipment assembly include assembly process cards, BOMs, torque records, gap adjustment records, commissioning reports, factory inspection sheets, and packing lists. Data sources include PLM/ERP, part inspection data, assembly tool collection, test benches, MES, and customer technical agreements. Updates are performed independently per device and per order. Fields with units include material codes, torque values (N·m), gap values (mm), and similar entries, organized by order dimension.
What constraints do these characteristics impose on knowledge base retrieval and recall
The multi-variety assembly per order characteristic requires retrieval to accurately match order-specific data for the current tender project, to avoid retrieving redundant documents across orders. Unit-bearing fields such as torque and gap require matching numerical ranges, which raises higher requirements for the semantic accuracy of retrieval. Frequently updated single-order documents require the retrieval scope to be limited to the dataset for the corresponding order, while also enabling rapid recall of core documents such as BOMs and process cards to support parameter verification and quotation preparation for tender documents.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the parameter precision requirements for general mechanical equipment assembly, filters low-relevance assembly process and BOM documents |
limit | Top 20 entries | Covers the number of assembly documents under 15 per single order, avoids exceeding the context carrying limit |
usingReRank | Enabled | Improves the sorting accuracy of precise numerical fields such as torque and gap, optimizes the relevance of retrieval results |
searchMode | embedding | Adapts to the matching requirements of multi-field and multi-type assembly documents |
datasetSearchUsingExtensionQuery | Enabled | Supplements technical agreement keywords related to tenders, enhances the recall precision of assembly parameters |
collectionIds | Dataset IDs bound to orders | Limits the retrieval scope, only recalls documents associated with the current tender project |
Three common mistakes
- Enabling
usingReRankduring online recall testing but seeing no effect. The returned results use the same sorting as the original semantic retrieval results, with no rearranged adjustments. The cause is that the re-ranking model is only enabled in global settings, and the function is not synchronously enabled in the index configuration of the corresponding knowledge base. - Setting
limitto 3000, and the context fields received by the large model are empty or have a length of 0. The cause is that the number of single recall entries exceeds the platform's default context carrying limit, triggering the truncation and discard logic, resulting in no valid content being passed. - Retrieval results include assembly documents from other orders that are unrelated to the current tender project. The cause is that
collectionIdsis not specified to limit the retrieval dataset scope, resulting in redundant assembly data being recalled across orders.
How to confirm the configuration is correct
- Enter the knowledge base retrieval test page of FastGPT v4.17.0, input "Assembly torque requirements for XX model reducer", check whether the sorting of the returned results matches the torque numerical matching degree, and confirm that
usingReRanktakes effect. - Adjust
limitto 20, verify that the number of context entries received by the large model matches the set value, and confirm that the truncation logic is not triggered. - Bind the
collectionIdsof the specified order, and only documents associated with that order will be returned during retrieval, confirming that the scope limitation takes effect. - Enable
datasetSearchUsingExtensionQuery, input "Assembly process parameters for tender use", and verify that the recall results include documents related to customer technical agreements.
The values provided here are common starting points and should be measured against the reader's own samples.