Skip to content

VERA Documentation

Step-by-step guides for running VERA, loading knowledge into it, and using it from an application or an AI agent.

Guides

  1. Getting started - run the full stack locally and confirm it works.
  2. Loading knowledge - the end-to-end flow to feed a repository, Markdown files, or Confluence into memory, and how VERA turns them into verified facts.
  3. Using VERA (HTTP API) - search, multi-hop explore, point-in-time queries, and retraction, with curl examples.
  4. Connecting an AI agent (MCP) - what the MCP server exposes and how a client connects to it.
  5. Integrate VERA with coding tools - one fast-path prompt and separate setup pages for Claude Code, Codex, and OpenCode.
  6. Agent integration GUIDE - the versioned, normative contract a coding runtime follows to wire VERA into an agent safely, with tested adapters for Claude Code, Codex, and OpenCode, plus a portable project setup skill that requires only the VERA API and MCP endpoint locations.
  7. Deployment - Docker Compose, Kubernetes, and the Helm chart, plus the operational commands.
  8. Architecture and algorithms - how VERA is built and the methods it uses: scoring, ranking, embeddings, the temporal model, and more.

The Mental Model in One Minute

  • Knowledge does not go straight into the graph. A connector pulls records from a source (a repo, a Markdown folder, Confluence), an extractor turns them into (subject, predicate, object) claims, and curation decides, by the source's trust tier, whether to publish them.
  • A published fact is stored in PostgreSQL (the source of truth) and projected into the Neo4j knowledge graph. It carries provenance: where it came from, how trusted it is, when it was true.
  • You read memory back through the HTTP API or the MCP server. Every read is scoped to the caller's tenant; every hit carries its provenance.
  • Content in many languages works: full-text search is language-agnostic and entity names keep their diacritics, so Vietnamese (and other non-English) knowledge indexes, matches, and deduplicates correctly.

For the architecture, algorithms (ranking, embeddings, scoring), and design rationale, see Architecture and algorithms.

Prerequisites

  • Docker (for the local infrastructure: Postgres, Neo4j, Valkey, MinIO).
  • The conda env vera (Python 3.11+) for the CLI commands.
  • An LLM/embedding key if you want to extract knowledge from free text (Markdown, Confluence, Git commits): OpenAI, or Voyage for embeddings and reranking. Structured sources (CMDB triples) work without one. See loading-knowledge.md.