What this category of data looks like
Core data sources for this category include advertising platforms, landing pages, product qualifications, and advertising law compliance clauses. Daily documents include advertising material text and format information, landing page screenshots, historical review opinions, and material modification records. Data updates occur daily or in real time. Document fields include material ID, violation type tags, rectification requirements, operation time, and adjustment content. Field attributes are flexibly adjusted based on document type, with no unified fixed units.
What constraints do these characteristics impose on knowledge base retrieval and recall
Differences in document types across multiple sources require retrieval to adapt to content in varied formats. Text-based review opinions and advertising law clauses require precise semantic matching. Landing page screenshots must be linked to text information of corresponding materials for consistency checks. Real-time or daily update frequencies require the knowledge base synchronization mechanism to pull the latest qualifications and clauses promptly, preventing recall of outdated content that reduces review accuracy. Material copy checking and landing page verification are core time-consuming tasks. Highly relevant compliance clauses and historical cases must be accurately recalled to avoid redundant results occupying context and lowering retrieval efficiency. The consequences of rejected advertisements require retrieval result confidence levels to align with audit decision requirements. Violation clauses that lead to missed or incorrect judgments must not appear.
How to configure settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Balances recall coverage and relevance when matching advertising law clauses and material violation points |
limit | Top 8 entries | Covers common violation types and historical review cases, controls context length for single-round retrieval |
searchMode | Semantic retrieval | Adapts to unstructured text matching needs for advertising material copy and review opinions |
usingReRank | Enabled | Performs secondary sorting on recall results by relevance to correct local deviations from a single similarity threshold |
datasetSearchUsingExtensionQuery | Enabled | Expands query semantic boundaries to match detailed segmented advertising platform placement rules |
collectionIds | Grouped by review scenario | Distinguishes between brand compliance and placement channel knowledge base collections, avoiding recall of irrelevant content across scenarios |
Three common misconfigurations
- Phenomenon: In simple workspace knowledge base applications, only 1 document is returned per retrieval, with a prompt showing "Knowledge base reference (1 entry)". Cause: The
limitparameter is not adjusted to a reasonable range, orsearchModeis configured for single-document matching mode. - Phenomenon: After importing a JSON-formatted review knowledge base, the knowledge base is empty when the application starts, and the dataset must be manually selected. Cause: The
collectionIdsparameter is not specified to bind the dataset for the corresponding business scenario, or the import configuration is not associated with the review scenario tag. - Phenomenon: An error is prompted when clicking test after configuring the model. Forcing ignore allows the application to run, but recall results are abnormal. Cause:
datasetSearchUsingExtensionQueryis not enabled, leading to insufficient query semantic expansion, making it impossible to match detailed segmented advertising platform placement rules.
How to confirm correct configuration
- Upload one piece of advertising material test text with clear violation points, and verify whether recalled review opinions and advertising law clauses match the material content.
- Adjust the
limitparameter, and check whether the number of returned retrieval results matches the preset value. - After enabling
usingReRank, check whether the order of recall results aligns with violation severity. - Upload one piece of real-time updated product qualification document, and verify whether retrieval results include the latest qualification information.
The values given 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.