grading

Document filtering uses AI-powered binary relevance scoring to automatically identify which retrieved documents answer your query. This Skill applies yes/no grading to each document, outputs a relevant boolean with reasoning, and enables downstream systems to retain only high-confidence results for answer generation.

Yes. Grading documents for relevance in retrieval-augmented generation pipelines removes noise from sources like ChromaDB or web search. When no relevant documents pass the grading threshold, the system can trigger query rewriting or retries, improving answer quality without manual review.

The grading Skill detects when zero documents meet relevance thresholds and signals downstream logic to initiate adaptive pipeline control—query rewriting or retry attempts—rather than generating answers from irrelevant sources.

Relevance grading applies binary classification (relevant/not relevant) to documents retrieved from ChromaDB, web search, or other sources. The Skill uses configurable AI models with structured outputs including relevance decision and reasoning explanation.

Document grading is config-driven through GRADING_MODEL in config.py, defaulting to Claude Haiku 3.5 with temperature 0 for consistent scoring. You can specify alternative models while maintaining structured result output.

Manual filtering is time-consuming and subjective; AI grading automatically scores all retrieved documents against query intent in seconds, identifies the most relevant subset for answer generation, and scales consistently across retrieval pipeline workflows.