Analytos Brain on Omnigraph — Submission POC

This repository is a submission-ready proof-of-concept for the Analytos Org Context Layer / Analytos Brain assessment.

It demonstrates the required loop:

Ingest → Extract → Human Review → Merge to main → Dashboard + MCP → Agents produce real work

What is included

• Omnigraph schema in omnigraph/schema/analytos.pg
• Omnigraph query templates in omnigraph/queries/*.gq
• Cedar policy sketch in omnigraph/policies/analytos.cedar
• Python ingestion pipeline that emits Omnigraph-compatible JSONL
• Idempotent deterministic IDs for nodes and edges
• Review/approval flow with branch diffs and merge attribution
• FastAPI dashboard for entity browsing, search, review, and recent changes
• MCP wrapper exposing governed graph reads to agents
• Content Agent and GTM Agent scripts
• Tests covering idempotency, governance, access control, and agent output
• Demo seed files matching the requested filenames

Note: The official private seed docs were not available in this chat, so this repo includes realistic demo seed files with the required filenames. If you have the official assessment seed files, replace files in seed-data/ and rerun the same commands.

Architecture

seed-data/*.md
  → pipeline.ingest
  → Gemini Flash extraction when GEMINI_API_KEY is set; deterministic fallback otherwise
  → runs/<run-id>/graph.jsonl
  → ingest/<run-id> branch
  → human review diff
  → approve merge to main
  → dashboard + MCP wrapper
  → content_agent.py and gtm_agent.py

The local graph store in .local_graph/state.json is used so the full demo can run without external services. The pipeline also emits Omnigraph-compatible JSONL and includes the real Omnigraph schema/query/policy files for a server-backed deployment.

Local setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Optional MCP dependencies:

npm install

Optional Omnigraph install:

curl -fsSL https://raw.githubusercontent.com/ModernRelay/omnigraph/main/scripts/install.sh | bash
omnigraph version

Optional: enable real Gemini LLM extraction

The repo runs without secrets, but for assessment compliance you should set a Gemini key locally or in hosting:

export GEMINI_API_KEY="your-rotated-google-ai-studio-key"
# optional
export GEMINI_MODEL="gemini-1.5-flash"

Never commit .env files or paste API keys into the repository.

Run the full local demo

bash scripts/demo.sh

This will:

1. Ingest all seed docs into ingest/demo-run.
2. Print the diff.
3. Approve and merge into main as human-reviewer.
4. Run Content Agent.
5. Run GTM Agent.

Manual workflow

1. Ingest seed data

python -m pipeline.ingest --input seed-data --run-id seed-run

Expected result:

branch: ingest/seed-run
status: pending_review
nodes/edges extracted

2. Review the branch diff

python -m pipeline.review diff ingest/seed-run

3. Approve and merge

python -m pipeline.review approve ingest/seed-run

Only human-reviewer can merge. ingest-agent cannot write directly to main.

4. Start dashboard

uvicorn dashboard.backend.main:app --reload --port 8000

Open:

http://localhost:8000

Dashboard pages:

• /entities — entity browser
• /search — search approved knowledge
• /review — pending branch review
• /recent — merge/commit history

5. Run Content Agent

python agents/content_agent.py "Stockly inventory forecasting"

The Content Agent:

• Uses only approved main-branch knowledge
• Cites graph node IDs
• Avoids EmailThread/internal-only nodes
• Shows access-control check

6. Run GTM Agent

python agents/gtm_agent.py "Stockly"

The GTM Agent returns:

• Target company profile
• Persona to contact
• Example companies
• Opening angle grounded in proof points
• Graph node citations

MCP usage

Install Node dependencies:

npm install

Run content-agent MCP wrapper:

ANALYTOS_ACTOR=content-agent node mcp/custom-wrapper/server.mjs

Run GTM-agent MCP wrapper:

ANALYTOS_ACTOR=gtm-agent node mcp/custom-wrapper/server.mjs

Claude Desktop-style configs are provided:

• mcp/content-agent-config.json
• mcp/gtm-agent-config.json

The MCP tool try_read_email_threads demonstrates policy behavior:

{
  "actor": "content-agent",
  "visible_count": 0,
  "denied_count": 2
}

Omnigraph-backed mode

The POC emits JSONL that follows the Omnigraph bulk load shape:

{"type":"Product","id":"product:stockly","data":{...}}
{"edge":"HAS_FEATURE","id":"edge:...","from":"product:stockly","to":"feature:stockly:demand-forecasting","data":{...}}

Initialize a real graph:

mkdir -p data
omnigraph init --schema omnigraph/schema/analytos.pg data/analytos-brain.omni

Load an approved run branch into Omnigraph:

omnigraph branch create ingest/seed-run data/analytos-brain.omni
omnigraph load --data runs/seed-run/graph.jsonl --mode merge --branch ingest/seed-run data/analytos-brain.omni
omnigraph branch merge ingest/seed-run --into main data/analytos-brain.omni

For cluster/server deployments, adapt omnigraph/cluster.yaml, then use:

omnigraph cluster validate --config omnigraph/cluster.yaml
omnigraph cluster plan --config omnigraph/cluster.yaml
omnigraph cluster apply --config omnigraph/cluster.yaml
omnigraph-server --cluster omnigraph/cluster.yaml --bind 0.0.0.0:8080

Tests

pytest -q

Covered criteria:

• Idempotent re-ingestion
• No direct writes to main by ingest-agent
• Merge requires human-reviewer
• content-agent cannot read EmailThread nodes
• Content Agent has citations and no internal client leak
• GTM Agent produces a prospecting brief

Assessment criteria mapping

Criterion	Where implemented
Governance correctness	pipeline/graph_store.py, pipeline/review.py, tests
Extraction quality	pipeline/extract.py, structured entities/edges, source metadata
Agent output quality	agents/content_agent.py, agents/gtm_agent.py
Access control	omnigraph/policies/analytos.cedar, local policy in LocalGraphStore.can_read
Dashboard usability	dashboard/backend/main.py
Engineering hygiene	README, tests, reproducible scripts, clear repo structure

Known limitations

• Gemini Flash extraction is implemented and used when GEMINI_API_KEY is set. The deterministic fallback remains for reproducible tests and demos without credentials.
• The local graph store is a test/demo fallback. Production submission hosting should run Omnigraph server with cluster.yaml and the Cedar bundle.
• The included seed docs are demo fixtures because the private official seed docs were not uploaded here.