Provena

Most RAG pipelines log the final answer and nothing in between. When a retrieved chunk turns out to be outdated, or a tool response looks suspiciously different from last week, there is precious little evidence to reconstruct what the model actually saw. Provena sits inside the Python pipeline — at the retriever, the tool call, or the context assembly step — and attaches a cryptographic hash plus metadata to each piece of context before it reaches the model. The vendor states this requires three lines of Python to instrument the first source.

The differentiating feature is tamper-evidence at the context level, not the output level. Rather than logging that a response was generated, Provena records what data was present, where it came from, and whether it has been modified since it was first ingested. The docs describe a policy engine that can flag staleness or provenance violations against configurable rules, and a PostgreSQL backend for durable audit storage — which makes the trail queryable rather than just printable.

Provena fits Python-based pipelines where the team needs to hand an auditor a timestamped, hashed record of every context source that contributed to a decision — the EU AI Act compliance use case the README names explicitly. It breaks down when the requirement shifts from ‘prove what happened’ to ‘prevent the bad thing from happening’: the library is passive. It records; it does not intercept. Teams building hard enforcement gates — block this retrieval if the source is untrusted — add a separate middleware layer, at which point Provena becomes the audit record for a system that enforces elsewhere.

The repository describes six framework adapters, covering common Python agent and RAG frameworks, and an MCP server for tool-call-level provenance. Self-hosted PostgreSQL is the stated storage backend. There is no commercial hosted offering; deployment and maintenance are the team’s responsibility.