lchoquel

It looks good, lean and clean. The optional ontology is a nice surprise. I was hoping for an SDK like that to plug into Pipelex for building a Knowledge Graph as a source to our multi-step AI workflows, I'll give it a try!

Working on it.

Hi atombender, you have a point. The current runtime supports asynchronous operations with the limits of python asyncio and your CPU/RAM. But we designed our orchestration layer so that it could integrate easily with durable workflow platforms such as temporal.io which also bring the ability to spread the load over many workers, which will give you horizontal scaling.

Our integration with temporal.io has been working as a prototype in our lab for many months but we had to focus on the core features to nail down our language and make it as easy to use as possible. Now in the coming weeks we'll be able to focus more on the temporal-powered version, adding the missing features to make it practical, which will also make it usable with other workflow platforms.

Hi teunlao,

Currently, a failed step means you need to restart everything: we have started to add features that save intermediate states so you can recover from there but the recovery part is not finished.

Our main strategy is to make Pipelex easy to integrate with durable workflow orchestrators which are really good at solving that problem. For instance, we'll soon release a plugin integration with temporal.io.

We get this question of comparison to BAML often, so let me take the time to clarify the differences:

Positioning:

• Pipelex is an AI workflow orchestrator. You declare multi-step pipelines (.plx) with LLM calls, OCR/PDF extraction, image generation, custom Python/tool steps, conditionals, and parallelism, then run them via CLI or Python, FastAPI/Docker, MCP, or n8n node.

• BAML defines typed, single-call “AI functions” (primarily LLM interactions) and compiles them into SDKs you call from your app; you typically orchestrate multi-step flows and any non-LLM work in your own code.

Developer experience:

• Pipelex feels like writing a readable pipeline spec you can iterate on or auto-generate and then execute end-to-end. Non-LLM ops (OCR, PDF, image gen, API calls) are first-class pipes you compose with LLM steps.

• BAML feels like writing typed prompt functions with templates and getting a clean, generated client API. Non-LLM work (OCR, PDF parsing, external APIs) usually sits in your app code or tools around those generated functions.

Structured output:

• Not exclusive to BAML. Pipelex’s PipeLLM supports strongly typed structured generation (based on the instructor package). Define schemas directly in .plx (simple concept syntax) or in Python with Pydantic v2, including custom validators and arbitrary validation logic.

Where each fits:

• If you want to design and run full workflows (extract → analyze → generate, with branching/parallelism) and keep that logic declarative and shareable, Pipelex is a great fit.

• If you mainly need reliable, typed “AI functions” embedded in an existing codebase, Pipelex also works well (define structures in Pydantic and call steps from Python). BAML is likewise a good fit when you prefer generated SDKs and you’re comfortable orchestrating steps and non-LLM work in app code.

Why Pipelex’s higher abstraction matters:

• It captures business logic and domain know-how in a structured, unambiguous language. Instead of burying intent across scattered prompts and glue code, a .plx file centralizes the specification (data concepts, pre/post-conditions, step semantics). That makes workflows easier to review, version, audit, and hand off, by humans and by AI, while remaining extensible to new AI/software operations.

TL;DR:

• Pipelex = declarative workflow engine/standard for multi-step AI pipelines, with first-class non-LLM ops and strongly typed outputs.

• BAML = typed prompt/function language that generates SDKs; you assemble and extend around it in your app.

When the business logic becomes complex, our motto which is to break it down into smaller problems, until they are manageable.

Regarding large docs, I know what you mean and we've been there: at one point we considered focusing on that feature, build a kind of agentic system to master large docs. At the time in 2023-2024, everyone was relying on vector store RAG and embeddings to solve that problem but we always thought that solution wouldn't bring enough reliability. We wanted the LLM to actually read the whole docs, in order to know what's in there and where. The idea was to read and "take notes" according to different aspects of what it's reading, and synthesize hierarchically from bottom to top. So, that part of the work could be done once and the structured "notes" could be exploited for each use case, pretty efficiently thanks to the zoom-in/out provided by the architecture. We went pretty far in building that solution, which we called "Pyramid RAG" internally, and the aim was to build AI workflows of top of that. But at some point, the building of the pyramid became pretty complex, and we realized we needed AI workflows to build the pyramid itself, and we needed to to do it right. That's when we decided to focus on what would become Pipelex. Now, in our new setting, solving the "large docs" problem is a large problem we can break down into a collection of Pipelex workflows. So it's part of our plan to provide those building blocks, and we hope to involve the community as much as possible, to cover all the possible applications.

Hi novoreorx, the biggest difference is that in Pipelex workflows, we express the logic in our high-level language, rather than in Python or typescript. This makes it easier to collaborate between tech and non-tech people, like domain experts. It's also great for collaboration with AI: the declarative approach and the high level of abstraction means that LLMs have all the context to understand what's going on in the workflow. Also, we don't use Pipelex for "autonomous agents": our workflows are meant to be repeatable and deterministic, like a tool. Like a tool that can be used by agents though our MCP server.

Hi Clafferty, Providing an MCP server was a no brainer and we have the first version available. But you're right, using "MCP as a client" is a question we have started asking ourselves too. But we haven't had the time to experiment yet, so, no definitive answer. For now, we have a type of pipe called PipeFunc which can call a python function and so possibly any kind of tool under the hood. But that is really a makeshift solution. And we are eager to get your point of view and discuss with the community to get it right.

Many companies are working on evals and we will have a strategy for integration with Pipelex. What we have already is modularity: you can test each pipe separately or test a whole workflow, that is pretty convenient. Better yet, we have the "conceptual" level of abstraction: the code is the documentation. So you don't need any additional work to explain to an eval system what we were expecting at each workflow step: it's already written into it. We even plan to have an option (typically for debug mode) that checks every input and every output complies semantically with what was intended and expected.

Thanks a lot for your feedback! It's a lot of work so greatly appreciated.

Yes! Clearly the introduction of LLMs into the mix raises the problem of throwing dice. The point of view we chose is: how to orchestrate the collaboration between AI, Software and people? With our aim to have repeatable workflows, this drove us away from building autonomous agents and towards a place where the software is in command of the orchestration. Then the Humans and AI can discuss "what you want to do" and have software run it and use AI where it's needed.