What Data for This Category Looks Like
The core entities for this category are power battery pack, energy storage cabinet, and BMS system manufacturers. Daily documents include grouping rules, structured BOM lists, welding and torque records, BMS flashing records, insulation and voltage withstand reports, EOL test reports, and system configuration lists. Data sources cover cell capacity grading data, MES systems, welding equipment, torque tools, BMS software version repositories, and EOL test benches. Data is synchronized once for each produced PACK or energy storage cabinet, and corresponding documents are updated each time the BMS software version changes. Document fields include cell capacity grading parameters, torque values, BMS version numbers, insulation test data, and more. No unified fixed format exists, and details vary based on specific business scenarios.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
Data sources are scattered across multiple types of equipment and systems. Retrieval tools must support cross-data-source document integration and incremental synchronization. Update frequency is high. Maintain knowledge base real-time performance to avoid recalling outdated BMS versions or eliminated grouping rules. Document types are diverse, and fields differ across documents. Adapt retrieval systems to multiple formats including structured BOMs, semi-structured test reports, and unstructured technical rules. Bidding scenarios require strict content accuracy. Filter irrelevant data tightly to avoid bidding errors caused by incorrect recalled information.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches professional technical parameters and rules for this category, avoids recalling irrelevant historical test data |
limit | Top 8 entries | Covers multi-dimensional technical parameters required for bidding, while controlling context length to avoid exceeding limits |
searchMode | embedding | Documents for this category mostly involve professional technical logic. Semantic retrieval more accurately matches the technical requirements of user queries |
usingReRank | Enabled | Re-ranks recalled documents by relevance, prioritizing the latest BMS versions and grouping rules |
datasetSearchUsingExtensionQuery | Enabled | Completes omitted professional terms in user queries, covering relevant data sources such as cell capacity grading and EOL tests |
collectionIds | Bind exclusive document collections for the current bidding project | Only calls documents such as BOMs and grouping rules for the current project, avoiding recalling outdated data from other projects |
Three Common Misconfigurations
- Symptom: Some queries in conversations do not trigger knowledge base retrieval, and generic answers are returned. Cause: Forced retrieval trigger conditions are not configured, or non-knowledge base answer results from previous conversations are cached.
- Symptom: After multiple rounds of queries in the same conversation window, knowledge base retrieval takes more than 60 seconds, while results return quickly in new windows. Cause: Historical context length is not limited, and accumulated context data after multiple rounds increases retrieval processing load.
- Symptom: Recalled documents include irrelevant cell raw material data under mixed retrieval mode, leading to insufficient answer accuracy. Cause: A matching mode tailored to professional technical documents for this category is not selected. Keyword matching in mixed retrieval introduces interference from non-target fields.
How to Verify Proper Configuration
- Manually enter a query containing cell capacity grading parameters or BMS version numbers, and check whether recalled results include corresponding business documents.
- View retrieval logs to confirm that the
similaritythreshold andlimitcount match the preset configuration. - Test more than 3 consecutive queries, and check whether retrieval latency remains stable with no obvious fluctuations.
- Switch between different retrieval modes, compare the relevance of recalled results, and confirm that the semantic retrieval mode’s matching effect fits business requirements.
The values provided here are common starting points, and should be evaluated against the reader's own samples.