Add a blueprint for Haystack Deep Research Agent

[oryx1729]

Description

Adds a blueprint demonstrating how to build a deep research agent using Haystack Framework that combines web search and Retrieval-Augmented Generation (RAG) using the NeMo-Agent-Toolkit.

By Submitting this PR I confirm:

• I am familiar with the Contributing Guidelines.
• We require that all contributors "sign-off" on their commits. This certifies that the contribution is your original work, or you have rights to submit it under the same license, or a compatible license.
  • Any contribution which contains commits that are not Signed-Off will not be accepted.
• When the PR is ready for review, new or existing tests cover these changes.
• When the PR is ready for review, the documentation is up to date with these changes.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[mdemoret-nv]

@oryx1729 Can you update your commits to contain the DCO? More info here: https://github.com/NVIDIA/NeMo-Agent-Toolkit/pull/461/checks?check_run_id=46372766560

[oryx1729]

Hi @mdemoret-nv, @mpangrazzi is on vacation and will be back next week. Is it be possible to start the review and have the DCO done when he's back?

[mdemoret-nv]

@oryx1729 No problem.

[mpangrazzi]

@mdemoret-nv Hi! Commits should be compliant with DCO now.

[mdemoret-nv]

Would be better to utilize the LLM top level object.

[mdemoret-nv]

We would want this to use the llm interface for specifying these models. Instead of referencing them directly, can we use the llms: top level object to specify this?

[willkill07]

Overall this looks good.

Please ensure that all "public" functions have full documentation and that all functions have type annotations. It really helps with clarity.

I can imagine in the future a complete haystack plugin with various features :)

[willkill07]

This currently does not use the builder at all

[willkill07]

No functions are registered with the toolkit directly. You are creating internal tools and building them within your specific workflow.

[willkill07]

I can envision a more robust separation of components, but the scope of that would be far too large. Instead, let's consider some enhancements to your existing configuration:

general:
  use_uvloop: true

llms:
  rag_llm:
    _type: nim
    model: nvidia/llama-3.3-nemotron-super-49b-v1
    api_key: ${NVIDIA_API_KEY}
  agent_llm:
    _type: nim
    model: nvidia/llama-3.3-nemotron-super-49b-v1
    api_key: ${NVIDIA_API_KEY}

workflow:
  _type: haystack_deep_research_agent
  max_agent_steps: 20
  search_top_k: 10
  rag_top_k: 15
  opensearch_url: http://localhost:9200
  data_dir: /data
  index_on_startup: true

By specifying the LLMs separately, they can automatically take advantage of other parameters like temperature. You can also still easily get the configuration by doing:

You can then get the configuration from rag_llm and instantiate the LLM by:

config = builder.get_llm_config("rag_llm")
generator = NVIDIAChatGenerator(**config.model_dump_json(exclude={"type"}))

You can do the same for agent_llm

[mpangrazzi]

@willkill07 @mdemoret-nv Thank you! I've just did another iteration following your suggestions.