One interface to every system of record.
Novexa sits between your data sources and your agents. It handles ingestion, chunking, embedding, access control, ranking and citation — and exposes all of it through a single stable API.
The problem we kept running into
Teams rarely fail at building an agent. They fail at giving that agent trustworthy context. A prototype wired to a vector store works in a demo, then collapses in production the moment it touches permissions, stale documents, contradictory versions, or a question that spans three systems at once.
We built Novexa so that the retrieval problem is solved once, centrally, and remains solved as data sources multiply.
Ingestion and connectors
Sources are attached declaratively. Each connector runs an initial backfill, then keeps the index current through change data capture, webhooks, or scheduled deltas — whichever the source supports best.
- Warehouses: Snowflake, BigQuery, Redshift, Databricks, Postgres, MySQL
- Object storage: S3, GCS, Azure Blob — with PDF, DOCX, PPTX and HTML parsing
- Knowledge tools: Confluence, Notion, Google Drive, SharePoint, GitHub
- Business systems: Salesforce, HubSpot, Zendesk, Jira, Linear
Retrieval pipeline
A query is normalized, expanded, and routed to the appropriate indices. Candidate chunks are retrieved by hybrid vector and lexical search, filtered against the caller's access scope, re-ranked by a configurable policy, then assembled into a context window with citations attached to every claim.
Every stage is versioned. You can pin an agent to a retrieval configuration the same way you pin a dependency, replay a golden question set against it, and diff the results before promoting a change.
Security and access control
Novexa treats the source system as the source of truth for permissions. Row-level and document-level ACLs are mirrored into the index and enforced at query time, so a retrieval never returns material the requesting identity could not open directly.
- SSO via SAML or OIDC; SCIM provisioning for seats and groups
- Per-workspace encryption keys, rotated on request
- Full audit trail of every retrieval, including the identity and the documents returned
- SOC 2 Type II report available under NDA; GDPR data processing addendum standard
Deployment options
The control plane is managed by us. The data plane is not. You choose where indexing and query serving actually run: our multi-tenant EU cluster, a single-tenant cluster in your own cloud account, or an air-gapped installation for regulated environments.
See it run against your own data.
Bring one data source to the call. We will index a slice of it live and query it together.