Leveraging ADK raw docs to feed LLMs to help coding

[devferrazZz]

Hello guys,

Every page in the doc has its raw version which is very LLM friendly. And I would guess they're conceived for this exactly purpose, maybe?

How could I leverage these pages to feed a LLM to help me code? Any scrapping-like or other approach?

Thanks

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: This is an excellent idea! The recommended approach is to use ADK's built-in tools for Retrieval-Augmented Generation (RAG) to create an agent that can answer questions based on the ADK documentation. ADK provides tools that connect to services like Vertex AI Search to make this process straightforward.

---

Hello! You've accurately identified a powerful way to use LLMs. Feeding documentation to an LLM to help with coding is a common and effective pattern, often called Retrieval-Augmented Generation (RAG). Your intuition that the raw markdown files are well-suited for this is spot on.

ADK is actually designed to facilitate this exact scenario. Instead of building a scraping and data processing pipeline from scratch, you can leverage ADK's built-in tool for RAG. [2]

Here’s a general approach you can take using ADK:

1. Prepare Your Data Source: As you suggested, cloning the adk-docs repository is a great first step. The raw markdown files are the perfect source.

2. Set Up a RAG Data Store: You'll need to ingest these documents into a service that can perform efficient vector search. A great option for this is Vertex AI Search. You can create a data store and upload the ADK documentation files to it.

3. Use ADK's RAG Tool: ADK provides a VertexAiRagRetrieval tool that allows an agent to query a Vertex AI Search data store. [1] You would equip your ADK agent with this tool.

4. Create and Instruct the Agent: You would then create an ADK agent and give it instructions to use the VertexAiRagRetrieval tool whenever it needs to answer a question about ADK. For a concrete example of how to implement a RAG agent with ADK, you can check out the Spanner RAG Agent sample. [3]

This approach allows the LLM to access and cite specific, up-to-date information from the documentation, making its coding assistance much more accurate and reliable.

For simpler, one-off questions, you could manually copy-paste the content of a specific raw markdown file into the LLM's context window, but the RAG approach is far more scalable and powerful.

---

References

[1] VertexAiRagRetrieval tool source code: https://github.com/google/adk-python/blob/main/src/google/adk/tools/retrieval/vertex_ai_rag_retrieval.py
[2] ADK Custom Tools Documentation (mentions RAG): https://google.github.io/adk-docs/tools-custom/
[3] Spanner RAG Agent Sample: https://github.com/google/adk-python/blob/main/contributing/samples/spanner_rag_agent/README.md

[devferrazZz]

It's an interesting reply but it's not quite what I'm looking for. 😄
@hangfei I'd love to get your input, mind taking a look?