Conduit

The knowledge graph engine. Graph-augmented retrieval for AI applications.

Conduit combines vector search with knowledge graph traversal to deliver retrieval that understands how concepts relate — not just how they're similar. Install a knowledge pack and query it in 3 lines of Python.

pip install 'conduit-ai[local]'

from conduit_ai import LocalConduit

conduit = LocalConduit("./my-knowledge-base")
conduit.install_pack("snowflake-2026.04.ckp")

results = conduit.search("How does Cortex Search work?")

Why Conduit

Every RAG system does vector search — find chunks similar to the query. Conduit does that plus graph traversal: when it finds a relevant concept, it walks relationships to discover connected concepts the query didn't mention.

A question about "Iceberg external tables" doesn't just find Iceberg docs — it walks edges to find related IAM configuration, Glue catalog setup, and cross-platform compatibility patterns. That's the difference between keyword-in-context retrieval and knowledge-aware retrieval.

Quick Start

Embedded (no server, no infrastructure)

pip install 'conduit-ai[local]'

from conduit_ai import LocalConduit

# Create a local knowledge base (DuckDB under the hood)
conduit = LocalConduit("./my-kb")

# Install a knowledge pack
conduit.install_pack("snowflake-2026.04.ckp")

# Search with graph-augmented retrieval
results = conduit.search("How does dynamic tables work?", limit=5)
for r in results:
    print(f"{r['score']:.3f} [{r['path']}] {r['title']}")

# Use as a LangChain retriever
retriever = conduit.as_retriever()
docs = retriever.invoke("Cortex Search vs Vector Search")

Server (full features)

git clone https://github.com/datakailabs/conduit.git
cd conduit
cp .env.example .env  # Fill in your values
docker compose up -d  # Start Postgres + ArangoDB
npm ci
npx tsc --skipLibCheck
node dist/src/server.js

# Ask a question
curl -X POST http://localhost:4000/api/v1/ask \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "How does Snowflake Cortex Search work?", "limit": 5}'

# Get raw context (no LLM synthesis)
curl -X POST http://localhost:4000/api/v1/context \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "Delta Live Tables patterns", "limit": 10, "format": "json"}'

Architecture

                    ┌──────────────────────────┐
                    │     Conduit Engine        │
                    │                          │
  Query ──────────► │  Vector Search (pgvector) │
                    │         +                │
                    │  Graph Traversal (Arango) │
                    │         +                │
                    │  LLM Synthesis (OpenAI)   │
                    │                          │
                    └──────────────────────────┘
                              │
              ┌───────────────┼───────────────┐
              ▼               ▼               ▼
         REST API        GraphQL         Python SDK

Retrieval pipeline:

1. Embed query → vector similarity search → top-K candidates
2. Walk knowledge graph 1-2 hops from candidates → discover related concepts
3. Merge vector results + graph discoveries → rerank by combined score
4. (Optional) Synthesize answer via LLM with graph-augmented context

Knowledge Packs

Knowledge packs are portable, versioned units of domain knowledge. Install them to make domains instantly queryable.

# Inspect a pack
conduit inspect snowflake-2026.04.ckp

# Install (full)
conduit install snowflake-2026.04.ckp

# Install (topic-scoped)
conduit install aws-2026.04.ckp --topics s3,iam,redshift

Seed packs available at datakailabs/knowledge-packs:

Pack	Zettels	Topics	Description
snowflake	5,634	cortex-search, iceberg, snowpark, ...	Snowflake platform
aws	3,466	s3, redshift, iam, glue, athena, ...	AWS services
databricks	4,173	unity-catalog, delta-lake, mlflow, ...	Databricks platform
genai	1,671	rag, embeddings, prompt-engineering, ...	GenAI patterns

Build your own packs from any content. See the Knowledge Pack Spec.

Integration

Python SDK

pip install conduit-ai                    # API client only
pip install 'conduit-ai[langchain]'       # + LangChain retriever
pip install 'conduit-ai[local]'           # + embedded engine (DuckDB)
pip install 'conduit-ai[all]'             # Everything

LangChain Retriever

from conduit_ai.retriever import ConduitRetriever

retriever = ConduitRetriever(
    api_key="ck_...",
    endpoint="http://localhost:4000",
)

# Drops into any chain
chain = {"context": retriever, "question": RunnablePassthrough()} | prompt | llm

MCP Server (Claude Code, Cursor, Pi, OpenCode)

pip install conduit-mcp

Gives AI coding agents access to your knowledge graph via MCP tools:

• conduit_ask — synthesized answers with sources
• conduit_context — raw knowledge units with graph relationships
• conduit_search — lightweight title/score search

See conduit-mcp for setup.

CLI

conduit ask "How does Cortex Search work?"
conduit inspect snowflake-2026.04.ckp
conduit install snowflake-2026.04.ckp --topics cortex-search,iceberg
conduit list

Deployment Modes

	Embedded	Server
Install	pip install 'conduit-ai[local]'	Docker Compose / k8s
Storage	DuckDB (single file)	PostgreSQL + ArangoDB
Graph	In-memory	ArangoDB (full AQL)
LLM	Bring your own	Built-in (OpenAI/Ollama)
Multi-user	No	Yes
API	Python only	REST + GraphQL
Scale	~50K zettels	~500K+
Use case	Notebooks, prototypes, CLI	Production, teams, chatbots

Tech Stack

• TypeScript / Node.js — server runtime
• PostgreSQL + pgvector — vector embeddings with HNSW indexes
• ArangoDB — knowledge graph with multi-hop traversal
• OpenAI — embeddings (text-embedding-3-small) and synthesis
• DuckDB — embedded mode storage + vector search
• GraphQL — query API for topology, search, and knowledge exploration

Contributing

See CONTRIBUTING.md. We welcome knowledge pack contributions, bug fixes, and feature proposals.

License

Conduit is licensed under AGPL-3.0. The Python SDK (conduit-py) is Apache-2.0.