synapsekit 1.5.5

Async-native Python framework for building LLM applications — RAG pipelines, tool-using agents, and graph workflows. Streaming-first, transparent API, 2 hard deps.

Documentation · Quickstart · API Reference · Changelog · Discord · Report a Bug

---

Build production LLM apps with 2 dependencies. Async-native RAG, Agents, and Graph workflows — no magic, no SaaS, no bloat.

"LangChain for people who hate LangChain."

SynapseKit is the minimal, async-first Python framework for LLM applications. 30 providers · 46 tools · 33 loaders · 9 vector stores. Every abstraction is plain Python you can read, debug, and extend. No hidden chains. No global state. No lock-in.

---

⚡ Async-native Every API is async/await first. Sync wrappers for scripts and notebooks. No event loop surprises.	🌊 Streaming-first Token-level streaming is the default, not an afterthought. Works across all providers.	🪶 Minimal footprint 2 hard dependencies: numpy + rank-bm25. Everything else is optional. Install only what you use.
🔌 One interface 30 LLM providers and 9 vector stores behind the same API. Swap without rewriting.	🧩 Composable RAG pipelines, agents, and graph nodes are interchangeable. Wrap anything as anything.	🔍 Transparent No hidden chains. Every step is plain Python you can read and override.

---

SynapseKit vs LangChain vs LlamaIndex

	SynapseKit	LangChain	LlamaIndex
Hard dependencies	2	50+	20+
Install size	~5 MB	~200 MB+	~100 MB+
Async-native	✅ Default	⚠️ Partial	⚠️ Partial
Cost tracking	✅ Built-in	❌ LangSmith (SaaS)	❌ No
Evaluation	✅ CLI + GitHub Action	❌ LangSmith (SaaS)	✅ Built-in
Graph workflows	✅ Built-in	✅ LangGraph (separate pkg)	❌ No
LLM providers	30	38+	20+
Stack traces	Your code	Framework internals	Framework internals

LangChain has more raw integrations and more tutorials. That's not what SynapseKit is optimizing for. SynapseKit is optimizing for the engineer who needs to ship, debug, and maintain an LLM feature in production — where readable code, predictable async behavior, and no surprise SaaS bills actually matter.

---

Who is it for?

SynapseKit is for Python developers who want to ship LLM features without fighting their framework.

• Burned LangChain users — hit a wall with debugging, dependency hell, or version churn and want full control back
• Async backend engineers — building FastAPI services where LangChain's sync-first model feels bolted on
• Cost-conscious teams — startups and teams who don't want a LangSmith subscription for basic observability
• ML engineers — building RAG or agent pipelines who need full control over retrieval, prompting, and tool use

---

What it covers

🗂 RAG Pipelines Retrieval-augmented generation with streaming, BM25 reranking, conversation memory, and token tracing. Load from PDFs, URLs, CSVs, HTML, directories, and more.	🤖 Agents ReAct loop (any LLM) and native function calling (OpenAI / Anthropic / Gemini / Mistral). 43 built-in tools including calculator, Python REPL, web search, SQL, HTTP, shell, Twilio, arxiv, pubmed, wolfram, wikipedia, and more. Fully extensible.
🔀 Graph Workflows DAG-based async pipelines. Nodes run in waves — parallel nodes execute concurrently. Conditional routing, typed state with reducers, fan-out/fan-in, SSE streaming, event callbacks, human-in-the-loop, checkpointing, and Mermaid export.	🧠 LLM Providers OpenAI, Anthropic, Ollama, Gemini, Cohere, Mistral, Bedrock, Azure OpenAI, Groq, DeepSeek, OpenRouter, Together, Fireworks, Cerebras, Cloudflare, Moonshot, Perplexity, Vertex AI, Zhipu, AI21 Labs, Databricks, Baidu ERNIE, llama.cpp, Minimax, Aleph Alpha, Hugging Face, SambaNova — all behind one interface. Auto-detected from the model name. Swap without rewriting.
🗄 Vector Stores InMemory (built-in, .npz persistence), ChromaDB, FAISS, Qdrant, Pinecone, Weaviate, PGVector, Milvus, LanceDB. One interface for all 9 backends.	🔧 Utilities Output parsers (JSON, Pydantic, List), prompt templates (standard, chat, few-shot), token tracing with cost estimation.
🧪 EvalCI — LLM Quality Gates GitHub Action that runs @eval_case suites on every PR and blocks merge if quality drops. No infrastructure, 2-minute setup. Score, cost, and latency tracked per case. Works with any LLM provider. → GitHub Marketplace · Docs

---

Install

pip

pip install synapsekit[openai]       # OpenAI
pip install synapsekit[anthropic]    # Anthropic
pip install synapsekit[ollama]       # Ollama (local)
pip install synapsekit[all]          # Everything

uv

uv add synapsekit[openai]
uv add synapsekit[all]

Poetry

poetry add synapsekit[openai]
poetry add "synapsekit[all]"

Full installation options → docs

---

Documentation

Everything you need to get started and go deep is in the docs.

🚀 Quickstart	Up and running in 5 minutes
🗂 RAG	Pipelines, loaders, retrieval, vector stores
🤖 Agents	ReAct, function calling, tools, executor
🔀 Graph Workflows	DAG pipelines, conditional routing, parallel execution
🧠 LLM Providers	All 30 providers with examples
🧪 EvalCI	LLM quality gates on every PR — GitHub Action
📖 API Reference	Full class and method reference

---

Development

git clone https://github.com/SynapseKit/SynapseKit
cd SynapseKit
uv sync --group dev
uv run pytest tests/ -q

---

Contributing

Contributions are welcome — bug reports, documentation fixes, new providers, new features.

Read CONTRIBUTING.md to get started. Look for issues tagged good first issue if you're new.

---

Community

• 💬 Discord — chat, help, show and tell
• 💬 Discussions — ask questions, share ideas
• 🧭 Discord roles draft — proposed roles and permissions for issue #389
• 🧭 Discord release webhook draft — automate release announcements for issue #390
• 🐛 Bug reports
• 💡 Feature requests
• 🔒 Security policy

---

Contributors

Nautiverse 💻 📖 🚧

Gordienko Andrey 💻

Deepak singh 💻

by22Jy 💻

Arjun Kundapur 💻

Harshit Gupta 📖

Dhruv Garg 💻

Adam Silva 💻

qorex 💻

Abhay Krishna 💻

AYUSH BHATT 💻

HARSH 📖

---

License

Apache 2.0