conduit-ai

Python SDK for Conduit — the knowledge graph engine.

Install

pip install conduit-ai                    # API client only
pip install 'conduit-ai[langchain]'       # + LangChain retriever
pip install 'conduit-ai[local]'           # + embedded engine (DuckDB, no server needed)
pip install 'conduit-ai[all]'             # Everything

Embedded Mode (no server required)

Run a knowledge graph locally — no Docker, no Postgres, no ArangoDB. Just DuckDB under the hood.

from conduit_ai import LocalConduit

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

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

# Formatted context for LLM prompts
context = conduit.context("Delta Live Tables patterns")

Topic-scoped installation — only load what you need:

conduit.install_pack("aws-2026.04.ckp", topics=["s3", "iam"])

LangChain Retriever (embedded)

retriever = conduit.as_retriever(limit=8)
docs = retriever.invoke("How do dynamic tables work?")

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

API Client (connect to a Conduit server)

from conduit_ai import ConduitClient

client = ConduitClient(api_key="ck_...", endpoint="http://localhost:4000")

# Ask a question (GraphRAG + LLM synthesis)
answer = client.ask("How does Snowflake Cortex Search work?")
print(answer.answer)
print(f"Sources: {len(answer.sources)}")

# Retrieve context without LLM
ctx = client.context("data pipeline best practices", limit=5)
for result in ctx.results:
    print(f"{result.title} ({result.score:.2f})")

Conversational follow-ups

import uuid

thread_id = str(uuid.uuid4())
answer1 = client.ask("What is Delta Live Tables?", thread_id=thread_id)
answer2 = client.ask("Can I use it with Cortex?", thread_id=thread_id)
# ^ automatically rewritten to: "Can I use Delta Live Tables with Snowflake Cortex?"

Streaming

async for token in client.aask_stream("Compare Databricks and Snowflake for ML"):
    print(token, end="", flush=True)

LangChain Retriever (server-backed)

from conduit_ai.retriever import ConduitRetriever

retriever = ConduitRetriever(
    api_key="ck_...",
    endpoint="http://localhost:4000",
    kai_id="kai_snowflake",   # Optional: scope to a Kai
    limit=8,
)

docs = retriever.invoke("How do I set up change data capture?")

CLI

Installed automatically with pip install conduit-ai:

# Inspect a knowledge pack
conduit inspect snowflake-2026.04.ckp

# Install a pack (full or topic-scoped)
conduit install snowflake-2026.04.ckp
conduit install aws-2026.04.ckp --topics s3,iam,redshift

# Dry run (preview without installing)
conduit install aws-2026.04.ckp --topics s3 --dry-run

# Ask a question
conduit ask "How does Cortex Search work?" --api-key ck_...

# List installed knowledge domains
conduit list

Knowledge Packs

Knowledge packs (.ckp files) are portable, versioned units of domain knowledge. Download seed packs from datakailabs/knowledge-packs:

Pack	Zettels	Description
snowflake-2026.04.ckp	5,634	Snowflake platform documentation
aws-2026.04.ckp	3,466	AWS services documentation
databricks-2026.04.ckp	4,173	Databricks platform documentation
genai-2026.04.ckp	1,671	Generative AI patterns and techniques

Two Modes

	Embedded (LocalConduit)	Server (ConduitClient)
Requires	pip install 'conduit-ai[local]'	Running Conduit server
Storage	DuckDB (single file)	PostgreSQL + ArangoDB
Graph	In-memory adjacency list	ArangoDB (full AQL)
LLM synthesis	Bring your own	Built-in
Multi-user	No	Yes
Scale	~50K zettels	~500K+
Best for	Notebooks, prototyping, CLI	Production, teams, chatbots

License

Apache-2.0