DEV Community

Cover image for Build governed, on-prem AI agents with VeritasGraph Studio (GraphRAG + citations)
BIBIN PRATHAP
BIBIN PRATHAP

Posted on

Build governed, on-prem AI agents with VeritasGraph Studio (GraphRAG + citations)

The problem: Most RAG demos answer "what does the doc say?". The hard question is the inverse: "who is violating the policy — and prove it."
Why a chatbot demo isn't enough for the enterprise (governance, audit, data residency)

Meet VeritasGraph Studio — one workspace, ten capabilities:

  • Agents — Compose role-focused agents (orchestrator, reviewer, assistant) and wire exactly which capabilities each one gets — Knowledge Graph, Tools, Memory, Guardrails, and Data logging — with a per-agent context budget.
  • Tools — Register callable tools (internal APIs, MCP servers, retrievers) with schema and auth, then run smoke tests before any agent can use them — so nothing ships un-validated.
  • Knowledge — Manage indexed corpora with hybrid search plus graph-linking, and tune chunk size and overlap for high attribution confidence.
  • Guardrails — Define hard-block and review policies — PII redaction, toxicity monitoring — enforced on every agent turn, with a live incident count for audit.
  • Memory — Give agents session (short-term) and long-term memory scopes, with hygiene checks that flag stale, duplicated, or conflicting memories.
  • Data — Connect operational data — SQL, files, events, object storage — and gate it with pre-index quality checks to reduce noisy retrieval.
  • Evaluation — Benchmark every agent for relevance, faithfulness, latency, and policy compliance, and track a rolling quality trend across runs — so you catch regressions before release.
  • Fine-tune — Queue domain-adaptation jobs on curated data slices, with enforced safety gates on every fine-tuned checkpoint.
  • Playground — Test any agent live against its local model and watch the full orchestration pipeline — guardrails in, memory, knowledge graph, context budget, tools, guardrails out, data log — with citations on every answer.
  • Knowledge Graph — Build a VeritasGraph knowledge graph from your documents and reason over it with verifiable citations and an explicit reasoning path.

The governed agent pipeline (guardrails in → memory → knowledge graph → context budget → tools → guardrails out → data log)

Citations + PROV-O lineage on every answer

Install & first run (local, Ollama):

pip install veritas-reason
pip install veritasgraph

Build a compliance agent end-to-end (create → wire tools/knowledge → evaluate → deploy)

Why enterprise teams pick it:

  • Governed by design — guardrails, evaluation, and data-quality gates are first-class, not bolted on.
  • Auditable — verifiable citations and PROV-O lineage on every answer for compliance and legal review.
  • Air-gapped — runs entirely on your infrastructure via Ollama; no data egress, no third-party API keys.
  • One workspace — build, test, evaluate, and govern agents in a single UI instead of stitching five tools together.

Wrap-up + repo link:https://github.com/bibinprathap/VeritasGraph
https://bibinprathap.github.io/VeritasGraph/studio/

Top comments (0)