Synthora

Next-generation deep research platform unifying three lineages:

• LangChain Open Deep Research — the research operating system: LangGraph nested graphs, a supervisor that plans and delegates, parallel researchers, tool loops, context compression.
• Stanford STORM / Co-STORM — the intelligence layer: multi-perspective expert personas, simulated collaborative discourse with a moderator surfacing unknown unknowns, hierarchical knowledge maps, outline-first cited writing.
• Local Deep Research — the platform layer: persistence, sessions, REST + WebSocket APIs, background jobs, pluggable search engines/strategies/LLM providers, self-hosting.

Research pipelines

Pipeline	Flow
fast_research	plan → search → summarize → answer
deep_research	plan → parallel research → perspectives → discourse → knowledge map → outline → cited sections → criticism → report
open_deep_research	plan → parallel research (supervisor loop only) → report
academic_research	literature search → citation verification → outline → synthesis → peer review → bibliography
autonomous_research	hypothesize → investigate → gap discovery → new research paths → knowledge base update → repeat (bounded)

Layout

apps/
  api/       FastAPI gateway (REST + WebSocket, optional auth)
  worker/    Queue consumer that executes LangGraph pipelines
  web/       React + Vite UI
packages/
  core/           Domain models, ports, events
  adapters/       LLM providers, search engines, search strategies, MCP bridge
  intelligence/   Perspectives, discourse, knowledge map, outline writer
  orchestration/  LangGraph pipelines + pipeline registry
  persistence/    SQLAlchemy models, repositories, Alembic migrations
  sdk/            Python client

Quickstart (development)

uv sync                                   # install workspace
uv run pytest                             # run tests
uv run uvicorn synthora.api.main:app --reload --port 8000
uv run python -m synthora.worker.main    # in another shell

LangGraph Studio: uv run langgraph dev (uses langgraph.json).

Self-hosting

docker compose up -d

Brings up: API (:8000), worker, web UI (:3000), Postgres, Redis, SearXNG. Default model profile is auto — resolves LLM roles from env keys, Ollama, or free fallback chains (llm_fallbacks via LiteLLM; set HF_TOKEN from hf auth login for Hugging Face inference).

Optional local LLM (Ollama profile):

docker compose --profile ollama -f docker-compose.yml -f docker-compose.ollama.yml up -d

Local-first preset (Ollama + SYNTHORA_MODEL_PROFILE=local):

docker compose -f docker-compose.yml -f docker-compose.local.yml --profile ollama up -d

The Ollama overlay sets OLLAMA_BASE_URL=http://ollama:11434 on API/worker. Embeddings follow profile (hash by default; Ollama when local profile active).

Key environment variables (see .env.example): SYNTHORA_MODEL_PROFILE (auto|local|cloud|free), SYNTHORA_DATABASE_URL, SYNTHORA_REDIS_URL, SYNTHORA_AUTH_MODE (none|session), OPENAI_API_KEY / OPENAI_BASE_URL, TAVILY_API_KEY, SEARXNG_URL, SYNTHORA_EXTENDED_ENGINES.

Docs

• docs/architecture.md — layer contracts and data flow
• docs/pipelines.md — pipeline graph designs
• docs/feature-parity.md — parity matrix vs the three source projects
• docs/eval-storm.md — STORM FreshWiki evaluation harness