Overview

When companies build internal AI tools, they pull data from databases, Confluence, Salesforce, and streaming topics — and feed it directly into vector databases or LLM context windows. The moment that happens, every access control from the source system disappears.

HR documents, salary records, customer PII, and healthcare records are flattened into vector embeddings with zero permissions, queryable by anyone with access to the AI tool.

Existing tooling does not solve this:

• SQL Access Controls (e.g. Immuta, Privacera) only monitor data warehouses, not streaming ingestion pipelines.
• LLM Guardrails filter user prompts at inference time, but sensitive data is already embedded in the vector DB by then.
• Traditional DLP is blind to high-throughput internal microservice data flows.
• Data Catalogs document lineage, but do not provide runtime enforcement.

Auvis sits at the pipeline layer — between your source data and your AI infrastructure — to enforce policies in real-time.

How Auvis Works

Auvis operates in two primary modes:

1. In-Flight Reverse Proxy Mode

Auvis sits between client applications and backend microservices or LLM APIs. When a request arrives:

1. Identity Extraction: Extracts user identity and role from headers (X-User-Role, X-User-ID), JWT bearer tokens, or Identity Providers (IdP).
2. Compile-Time Graph Lookup: Matches the request against pre-compiled policy rules in microseconds.
3. Stream Transformation: Masks emails, removes SSNs, pseudonymizes IDs with SHA-256 salts, or truncates strings.
4. Destination Fan-Out: Synchronously writes governed records and vector embeddings to downstream sinks (pgvector, Pinecone, Weaviate).
5. Non-Blocking Audit: Appends an immutable JSONL audit record with compliance tags (HIPAA, GDPR).

2. Continuous Streaming Pipeline Mode

Auvis runs as a native Kafka/Redpanda consumer group worker. It reads raw streaming records, applies governance policies, and routes sanitized records into storage sinks before downstream consumers or RAG retrieval pipelines can access them.

Core Guarantees

• Fail-Closed Governance: If an evaluation fails or an unauthorized user requests protected fields without a policy, data is withheld or masked by default.
• Zero Client Code Changes: Applications send standard HTTP requests through Auvis; no custom SDKs or client library migrations are required.
• Decoupled Sink Availability: Independent 3-state circuit breakers protect storage sinks. If a vector database is down, client responses are delivered without blocking.
• Zero-Downtime Hot Reload: Edit any YAML policy file in policies/ and changes take effect immediately across all worker goroutines without restarting the gateway.

Next Steps

• Follow the 5-Minute Quickstart to spin up Auvis locally with your first policy.
• Learn how to write fine-grained governance rules in the Policy DSL Reference.
• Check out the Configuration Reference for auvis.yaml options.