Recently I'm working on a project that let the users to analyze and rate the bid documents based on one tender document. The general logic is quite alike to the cv-job_offer example in Pipelex. The challenge I encountered is that both tender and bid documents are very large, it's impossible to load even a single one in LLM context, not to say both. Therefore I have to design a workflow to extract structured informations and evaluation criteria from the tender doc into predefined data models, and then dispatch multiple tasks to complete evaluation of these data models against each bid doc.
I'm wondering if this kind of scenario (essentially it's just the input document is too big) is possible to be handled in Pipelex. In my understanding DSL is good for it's high-level abstraction and easy to understand, but lacks flexibility and the power is restricted. How can the users of Pipelex iterate the pipelines to fulfill the complex need when the business logic became complex inevitably?
Declarative workflows is such a good idea, fantastic, and I love the AI first principles where pipeline creation and editing the pipeline can be done with AI too.
The declarative style keeps the workflow detail at a high enough level to iterate super quick - love that. More important to me is that it’s structured and seems like it would be more testable (I see validation in your docs).
Zooming in to the pipe/agent steps I can’t quite see if you can leverage MCP as a client and make tool calls? Can you confirm? If not what’s your solution for working with APIs in the middle of your pipeline?
Also a quick question, declarative workflows won’t solve the fact that LLMs output is always non deterministic, and so we can’t always be guaranteed the output from prior steps will be correct. What tools or techniques are you using/recommending to measure the reliability of the output from prior steps? I’m thinking of how might you measure at a step level to help you prioritise which prompts need refinements or optimisations? Is this a problem you expect to own in Pipex or one to be solved elsewhere?
Great job guys, your approach looks like the right way to solve this problem and add some reliability to this space. Thanks for sharing!
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.
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.
Sorry, I guess I'm not fully understanding what this is exactly. Would you describe this as a low-code/no-code agent generator? So if you can define requirements via a pipelex "config" file, Pipelex will generate a python-based agent?
Hi RoyTyrell,
I guess you could call it low-code, a new kind of no-code where we have natural language in the mix.
But no, Pipelex does not generate a python-based agent: the pipelex script is interpreted at runtime.
Declarative DSL is a really interesting approach, especially since you’re exposing it directly to the users. There are some applications where throwing the dice in production by having LLM as part of the runtime is not an option.
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.
Very cool declarative + agent-first is the right direction. Love the “Dockerfile for AI reasoning” analogy. Excited to try composing Pipelex with Codiris workflows.
I'm wondering if this kind of scenario (essentially it's just the input document is too big) is possible to be handled in Pipelex. In my understanding DSL is good for it's high-level abstraction and easy to understand, but lacks flexibility and the power is restricted. How can the users of Pipelex iterate the pipelines to fulfill the complex need when the business logic became complex inevitably?
The declarative style keeps the workflow detail at a high enough level to iterate super quick - love that. More important to me is that it’s structured and seems like it would be more testable (I see validation in your docs).
Zooming in to the pipe/agent steps I can’t quite see if you can leverage MCP as a client and make tool calls? Can you confirm? If not what’s your solution for working with APIs in the middle of your pipeline?
Also a quick question, declarative workflows won’t solve the fact that LLMs output is always non deterministic, and so we can’t always be guaranteed the output from prior steps will be correct. What tools or techniques are you using/recommending to measure the reliability of the output from prior steps? I’m thinking of how might you measure at a step level to help you prioritise which prompts need refinements or optimisations? Is this a problem you expect to own in Pipex or one to be solved elsewhere?
Great job guys, your approach looks like the right way to solve this problem and add some reliability to this space. Thanks for sharing!
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.
[1]: https://github.com/boundaryml/baml
Waiting for partnership to propose to our users