Architecture

Ontora’s pipeline is built around three layers:

Conversation layer — voice or chat interviews conducted by an autonomous agent.
Knowledge layer — transcripts processed through chunking, embedding, and entity extraction into Postgres + Neo4j.
Insight layer — synthesis outputs (cartography, roadmap, personas) and a GraphRAG query endpoint over the campaign.

Pipeline

When a campaign completes a conversation, a job pipeline kicks off:

INTERVIEW_COMPLETED
  ├─► CHUNK_DOCUMENT  (sentence-boundary, 1000 chars, 200 overlap)
  │     └─► EMBED_CHUNKS  (text-embedding-3-small → pgvector)
  └─► EXTRACT_ENTITIES_RELATIONS  (gpt-4o-mini → JSON)
        └─► UPSERT_GRAPH  (Neo4j, deduped by stable keys)
              └─► SYNTHESIZE  (cartography + roadmap + personas)
                    └─► webhook: synthesis.completed

Every stage is idempotent and retried up to 3× with exponential backoff.

Multi-tenancy

Every record carries a workspace_id. API keys are workspace-scoped — there is no cross-workspace data access. See Workspaces.

What gets stored

Store	Holds	Used for
Postgres (pgvector)	Documents, chunks, embeddings, jobs, conversations	Vector search, transactional state
Neo4j	Entities (Person, Topic, Process), relations	Graph traversal during GraphRAG
Object storage	Raw transcripts, audio	Export endpoints

Three integration surfaces

The same data is exposed through three interfaces, all backed by the same workspace API key:

REST API — request/response for programmatic integration
MCP server — tool-calling interface for AI agents
CLI — terminal and CI-friendly wrapper around the REST API

Pick whichever fits your environment; mix them freely.

Get started

Core concepts

Guides

MCP server

CLI

Webhooks

Pipeline

Multi-tenancy

What gets stored

Three integration surfaces

​Pipeline

​Multi-tenancy

​What gets stored

​Three integration surfaces

Pipeline

Multi-tenancy

What gets stored

Three integration surfaces