Prove what your AI saw.
When an answer is challenged by an auditor, a regulator or a customer's security review, most teams cannot prove what their model was actually given. Logs show requests and responses; they do not prove the retrieved context was authentic, came from an approved source, or should have been allowed through. Traction Layer's provenance products treat the model's input as a control surface: signed at ingestion, verified at inference, governed by policy, recorded as evidence.
RAG Provenance Platform
The control and evidence layer for enterprise retrieval-augmented AI, covering vector, hybrid, GraphRAG and agentic retrieval.
The gap it closes.
AI guardrails inspect prompts but not the integrity of the knowledge pipeline that feeds them. Data-loss and network tools control data movement but cannot say whether a chunk was tampered with, whether it came from an approved corpus, or whether policy should have blocked it. Nothing in a standard stack produces evidence that the retrieved context was what it claimed to be.
What it does.
- 01Sign
Signs every retrieval unit at ingestion. Chunks, structured query results and graph subgraphs are signed with identity-bound certificates built on Sigstore primitives, compatible with enterprise transparency logs including Red Hat Trusted Artifact Signer.
- 02Verify
Verifies at inference. Before any retrieved object reaches the model, the digest is recomputed and the signature checked against the corpus trust policy. The decision is allow, warn or block. In closed mode the system fails closed: tampered, stale or wrongly signed content never reaches the model.
- 03Manifest
Emits an Assembly Manifest per request. A signed record of which retrieval units contributed, who signed them, what trust tier applied, what policy decision fired, and a hash linking the model's response back to the verified context.
- 04Ledger
Keeps an append-only provenance ledger. Signing events, verification outcomes, policy decisions and request metadata, hash-chained and indexed for forensic query. Reconstruct what the AI saw in minutes, not days.
- contributing retrieval units
- [retrieval units]
- signers
- [signer identities]
- trust tier
- [trust tier]
- policy decision
- [allow, warn or block]
- response hash
- [response hash]
Customer-resident by design.
Version 1 deploys inside your AWS account: your IAM, your KMS, your VPC, your object store, your logs. Traction Layer holds zero customer data: no prompts, no responses, no chunk content, no manifests. That is provable by architecture, not promised by policy. Optional encrypted response storage uses your customer-managed keys.
Works with your stack.
- Vector backends: ChromaDB, OpenSearch, Pinecone, Qdrant
- Source systems: Amazon S3, SharePoint, Confluence, SQL databases, file shares, API feeds
- Identity: SAML, OIDC, AWS IAM Identity Center
- Observability: OpenTelemetry-native; exports to Splunk, Microsoft Sentinel, Google Chronicle or any SIEM
- Models: provider-agnostic; hosted model APIs, Amazon Bedrock and self-hosted engines including Traction Runtime
Built for.
CISOs and security architects facing AI risk reviews; GRC, audit and compliance teams who need evidence beyond application logs; AI platform leaders deploying RAG in regulated workflows; financial services, healthcare, government and regulated SaaS.
Input Context Provenance Service
Retrieval is one input among several. The Input Context Provenance Service extends the same model to the whole input surface: system prompts, retrieved chunks, tool outputs, conversation history, uploaded files and fetched URLs. A Context Assembler in front of any inference backend (Amazon Bedrock, Amazon SageMaker, self-hosted vLLM or SGLang, any OpenAI-compatible endpoint) verifies the provenance of every component, fails closed on any failure, and writes a signed in-toto Assembly Manifest to write-once object storage. AWS-native, single account, no customer data leaves the boundary. The reference architecture is published; RAG provenance ships first and grows into full input-context coverage.
Evidence for EU AI Act Article 12 and the frameworks your assessors use
Article 12 of the EU AI Act requires high-risk AI systems to keep logs that allow their operation to be traced. An Assembly Manifest per request and a hash-chained ledger are exactly the kind of record that requirement describes. We say this carefully: our products generate the evidence; they do not certify your system. Control mappings to NIST AI RMF and ISO/IEC 42001 are provided as alignment, not certification, and Traction Layer does not hold SOC 2 or ISO/IEC 42001 today.
Traction Layer AI aligns its controls and audit records to NIST AI RMF, ISO/IEC 42001 and EU AI Act Article 12. Alignment is not certification. Traction Layer AI does not hold SOC 2 or ISO/IEC 42001 certification today; both are on our certification roadmap.
Start the conversation
See how provenance fits your retrieval pipeline. The first step is a working session with an engineer on your architecture, followed by an early-access deployment in your AWS account.