EPIC - Python feature parity with Typescript

[mmurad2]

Description

This Epic focuses on achieving feature and functionality parity between our Python framework and the existing TypeScript framework. This Epic is primarily for developers working with the Python version of the BeeAI framework, ensuring they have the same tools and capabilities as TypeScript users.

Initiative / goal

This Epic supports the larger initiative of increasing adoption of the Python version of framework.

Hypothesis

Most impactful features to prioritize:

• CLI getting started experience
• Agents (+ examples)
• Workflows (+ examples)
• LLMs: ollama, watsonx, groq, openAI (rest can follow later)
• Templates
• Memory
• Tools: Code interpreter, custom tool, integrations with Langchain and MCP?
• Observability (dependency [add epic])

Acceptance criteria and must have scope

TBD

Stakeholders

Granite Bee (in python): @grahamwhiteuk @michael-desmond
Testing / feedback on dev experience: @ismaelfaro

Timeline

Goal is to complete full feature parity by end of Feb.
Most critical features to be shipped ahead by Feb 10 -- @grahamwhiteuk @michael-desmond to provide requirements.

[mmurad2]

@michael-desmond @grahamwhiteuk I am assigning you the ticket while it is still in planning so you can share the features you need prioritized in python to meet the Feb 12 deadline.

[vabarbosa]

also adding @ajbozarth as he will be assisting me with some of this work

[michael-desmond]

For the Feb 12 deadline my take is that the follwing features should be prioritized:

• Ollama + json structured outputs (OllamaChatLLM and JsonDriver in .ts world)
• Workflows
• Memory, Message primitives etc.

The reasoning here is that we want hackathon participants to be able to create interesting agents with minimal setup complexity. Ollama gives us an easy way to access granite 3.1, no need for cloud credentials etc. Workflows allow the definition of complex agents and multi-agent systems. The ollama structured output (json) is necessary as it allows you to produce tool parameters, and then you can just call the tool yourself. So participants could just use the langchain tools in the meantime.

[michael-desmond]

Example of calling a tool directly. We could provide examples of this pattern to get people familiar. ETA this is a workflow step i.e. a node in an agent graph.

.addStep("search", async (state) => {
  
  const driver = new JsonDriver(getChatLLM());
  const { parsed } = await driver.generate(
    z.object({
      search_query: z.string(),
    }),
    [
      BaseMessage.of({
        role: Role.SYSTEM,
        text: `You are an expert at creating web search queries.`,
      }),
      BaseMessage.of({
        role: Role.USER,
        text: SEARCH_QUERY_PROMPT.render({ task: state.input }),
      }),
    ],
  );

  const searchResults: SearxngToolOutput = await new SearxngTool().run({
    query: parsed.search_query,
  });

  return { update: { results: searchResults }, next: "synthesize" };
})

[Tomas2D]

@michael-desmond drivers are going to be dropped, see #298.
However, native tool calling will be available, and the provider's structured output will be available (which is more reliable).

[michael-desmond]

Thanks, structured output is the key feature.