Observability and measuring recall quality

[pablomdc]

Is there built-in metadata in recall results (scores, confidence, source tracing) to detect low-quality retrievals before passing them to the LLM, specifically when using auto_route=True? And is there are commended pattern to measure over time whether memory is actually improving agent decisions?

[VectorPeak]

Short answer: partially. recall(auto_route=True) is a high-level API, not a calibrated retrieval-quality API. You get routing/provenance metadata: source (session, graph, trace, graph_context), and graph results include fields like text, kind, search_type, optional dataset fields, metadata, and raw. For chunk/summary paths, metadata can include data_id, chunk_id, chunk_index, and document_name; lexical chunk raw may include scores. Completion-style paths generally should not be treated as having a universal confidence score. (docs.cognee.ai)

---

If my answer solved your problem, please mark it as answered the question — I'm here to help, and honestly I'm also collecting Galaxy Brain badges along the way 😆

[pablomdc]

Looking for two things:

1. Built-in quality signal on recall results — is there any score, confidence, or relevance field returned by recall() (especially with auto_route=True) that can be used to filter low-quality results before they reach the LLM?
2. Best practices for measuring retrieval quality — how are others evaluating whether recall is actually returning relevant context? Looking for patterns around precision/recall metrics, threshold tuning, or any recommended eval setup.