Orchestration: LangGraph and its alternatives
| Solution | Publisher | Origin | Open source | Self-hostable | Pricing model | Note |
|---|---|---|---|---|---|---|
| LangGraph (to keep) | American publisher | International (open source) | Yes (MIT) | Yes | Free | Self-host; avoid the managed cloud platform |
| Haystack (deepset) | Germany | EU | Yes | Yes | Free | |
| LlamaIndex | American publisher | International (open source) | Yes | Yes | Free |
LangGraph is an MIT-licensed library: the code runs wherever you want. The dependency appears when you adopt the managed cloud platform that comes with it. I use it self-hosted, with a state store (checkpointer) on a PostgreSQL database I control. Haystack, developed by deepset in Berlin, is the most complete European alternative.
Agent tracing: replacing LangSmith
| Solution | Country | Origin | Open source | Self-hostable | Pricing model |
|---|---|---|---|---|---|
| Langfuse | Germany | EU | Yes | Yes | Free / freemium cloud |
Tracing (traces, costs, evaluations) is the tool that sees all the data pass by: prompts, answers, documents. It is the one to self-host first. Langfuse allows it, with direct integration into LangGraph and LiteLLM.
Multi-provider gateway
| Solution | Publisher | Origin | Open source | Self-hostable | Pricing model | Note |
|---|---|---|---|---|---|---|
| LiteLLM | American publisher | International (open source) | Yes | Yes | Free | Switch between providers |
LiteLLM exposes a single interface in front of every provider: Bedrock, Infomaniak, Mistral, Scaleway, a local vLLM. The application only knows the gateway; you route by rule (sensitive data to a Swiss API, the rest to the most suitable model) and change provider without touching the code.
Vector database: replacing Pinecone
| Solution | Country | Origin | Open source | Self-hostable | Pricing model |
|---|---|---|---|---|---|
| Qdrant | Germany | EU | Yes | Yes | Free / pay-as-you-go cloud |
| Weaviate | Netherlands | EU | Yes | Yes | Free / pay-as-you-go cloud |
A vector database holds your documents chunked and indexed: it is a copy of your knowledge. Qdrant (Berlin) and Weaviate (Amsterdam) are free software, host on any European cloud, and offer a managed cloud if you prefer. For modest volumes, PostgreSQL’s pgvector extension even avoids one more component.
The resulting architecture
- Application and orchestration (LangGraph or Haystack) hosted on a Swiss or European cloud.
- LiteLLM gateway in front of the model providers, with routing by sensitivity.
- Self-hosted Langfuse tracing, never in a third-party cloud for client data.
- Qdrant, Weaviate or pgvector vector database on the same infrastructure.
- State, rules, prompts and evaluation sets in your repository and your database, not at a provider.
I put this kind of migration in place with your teams: inventory, choices by data sensitivity, staged switch-over, operations. A first conversation is enough to know where to start.
Frequently asked questions
Do we have to drop LangGraph to be sovereign?
No. The library is free software (MIT) and self-hostable. What to avoid is the managed cloud platform that comes with it, and storing state at a third party.
Why self-host tracing first?
Because the tracing tool sees every prompt and every answer, hence all the data. Self-hosted Langfuse settles the question.
Qdrant or Weaviate?
Both are suitable. Qdrant is lighter to operate; Weaviate offers more built-in modules. For small volumes, pgvector in PostgreSQL is enough.