This document is written for FastGPT v4.17.0.
What this category’s data looks like
Core data for this category comes from ad platform backends, attribution systems, CRM lead databases, and landing page data. Update frequencies are mostly daily or real-time. Daily produced documents include delivery daily and weekly reports, single price adjustment records, material AB test sheets, and optimization review notes. Common fields include ad plan ID, consumption amount, conversion volume, material click-through rate, and more. Units include yuan, times, percentage, and others. Data volume changes dynamically with delivery scale. This data is primarily used for abnormal indicator locating, budget allocation, and material effect review.
Constraints on the knowledge base retrieval and recall link
This category has multi-source data with flexible update frequencies. This requires retrieval to support cross-data source associated recall, avoiding fragmented information. Daily documents have varied structures, ranging from structured price adjustment records to unstructured review notes. Retrieval must adapt to content splitting and matching across different formats. Real-time updated delivery data requires retrieval latency to stay under 10 seconds. Otherwise, real-time issue responses for on-duty operations cannot be supported. Errors in delivery performance adjustment directly impact business costs. The relevance and accuracy of recall results must match the precision requirements of business decisions.
How to set configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the information relevance requirements of delivery business, filters low-relevance redundant results |
limit | Top 6 entries | Controls the number of results returned per retrieval round, adapts to the quick reading needs of on-duty operations |
usingReRank | Enabled | Optimizes ranking effects for mixed structured and unstructured documents, improves recall accuracy for review-related content |
searchMode | hybrid | Combines keyword and semantic retrieval, adapts to mixed retrieval of structured data such as price adjustment records and unstructured data such as review notes |
collectionIds | Specify collection IDs related to ad delivery and attribution analysis | Focuses on category-specific data, filters non-delivery redundant knowledge base content |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval terms, covers common industry term variants used by ad optimizers |
Three common mistakes
- Phenomenon: Context truncation error occurs after retrieval results are passed to AI chat. Cause: Total token count of knowledge base content and user questions is not correctly calculated, exceeding the model's supported context window.
- Phenomenon: Number of returned retrieval results does not match the set
limit. Cause: Result count limit for the re-ranking stage is not synchronously configured, leading to adjusted result counts after re-ranking. - Phenomenon: Retrieval and recall results show no obvious change after optimizing configurations. Cause: The knowledge base node's problem optimization and the independent problem optimization node are confused. The former optimizes problem mappings within the knowledge base, while the latter optimizes user input query terms.
How to confirm configurations are correctly set
- Manually input a typical business question, such as "How to adjust budget when yesterday’s ad consumption exceeds the standard", and check if retrieval results include content related to price adjustment records and budget allocation.
- View retrieval logs, confirm that the
similarityscore of returned results falls within the preset 0.75–0.85 range. - Test real-time updated delivery data, confirm that retrieval results are returned within 10 seconds.
- Check the association relationship of knowledge base collections, confirm that
collectionIdsonly includes category-specific knowledge base content.
The values provided here are common starting points, and should be evaluated against applicable 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.