What the data for this category looks like
Core data for the IT asset inventory scenario comes from asset management systems, procurement systems, financial fixed asset modules, and terminal management platforms. Daily generated documents include asset ledgers, monthly inventory sheets, loan and return forms, and scrap application forms. Data fields include asset number, device name, specification model, user, deployment location, and status. Additional relevant fields are also included. Document structures exported from different systems vary. Data updates follow monthly, quarterly, or annual cycles. The total size of a single summary document increases as the inventory scope expands.
Constraints Imposed on Context and Token Processing
Differences in document formats across multiple sources require context processing to support varied field layouts. Context processing must prevent critical associated information such as asset numbers and users from being lost during text splitting. Low update frequency paired with large single-summary document volumes leads to a significant increase in context token consumption. Strict control over recall and truncation thresholds is required. Field priorities differ across document types. Core fields of asset ledgers must be prioritized for inclusion in context. Approval process information on scrap application forms only needs to be recalled in specific scenarios. The account-physical reconciliation business scenario requires context to retain association links between historical inventory records. Field breaks across documents must be avoided, and the scenario must be distinguished from configuration management CMDB maintenance scenarios. It focuses more on real-time associations of asset ownership, location, and user assignments.
Configuration Settings
These configuration settings are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | 12000–15000 token | Adapts to the total document token scale of IT asset inventory summary scenarios, prevents the model from truncating critical asset information |
quoteMaxToken | 4000–6000 token | Limits the token occupancy of a single recalled document, balances information completeness and context redundancy |
| Chunk overlap length | 100–150 characters | Retains contextual continuity for cross-chunk associated fields such as asset numbers, users, and locations |
| Single recall count | Top 8 entries | Matches the number of entries in single-batch data such as monthly inventory sheets and ledgers, avoids excessive irrelevant information interference |
| Chunk length | 800–1000 characters | Adapts to the field length of single records in asset ledgers, prevents single chunks from containing too many unrelated asset information |
Three Common Configuration Mistakes
- Phenomenon: The document format displayed in the context reference area is plain text, and the original document’s Markdown formatting is not retained. Cause: Document format retention configuration is not enabled, or format markers from the original document are lost during splitting.
- Phenomenon: The number of context reference entries displayed on the page does not match the actual number of entries sent to the model. Cause: Associated documents with duplicate asset numbers are not properly filtered during recall, or the context truncation threshold setting is inconsistent with the page display logic.
- Phenomenon: The token consumption statistics returned by the model deviate from the manually calculated document token count. Cause: Asset field units and special characters are not processed according to FastGPT’s token calculation rules, or token consumption from non-core fields is not excluded.
How to Verify Proper Configuration
- A standard asset ledger document is uploaded, and the format of the context reference area is verified to match the original document.
- The number of context entries displayed on the page is compared to the actual recall entries in the model call log, and the recall configuration is adjusted to align the two values.
- The token count of a single inventory sheet is manually counted, compared to the token calculation result from FastGPT, and the adaptability of the calculation rules is verified.
- The API is called with custom context parameters, and the model’s ability to correctly receive and use the passed context content is confirmed.
The values provided are common starting points and should be measured against the reader's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.