Exploring inspectable RAG pipelines on a fully local Ollama setup

Posted by HarinezumIgel@reddit | LocalLLaMA | View on Reddit | 3 comments

I’ve been working on RAG‑LCC (Local Corpus & Classification), an experimental, offline‑first RAG lab built around a fully local Ollama setup.

The goal isn’t to ship a production framework, but to experiment with and inspect RAG behavior—document routing, filtering stages, and retrieval trade‑offs—without hiding decisions inside a black box.

Current assumptions / constraints

Local‑only operation
Ollama is the only backend tested so far
No cloud dependencies
Tested on Windows 11 so far
Designed for experimentation, not production use

What I’m exploring

Classify‑then‑load document routing instead of indexing everything
Staged retrieval pipelines where each step is observable
Combining classical heuristics with embeddings and reranking

For interactive use, the project can optionally start a local OpenAI‑compatible listener so Open WebUI can act as a front‑end; the UI is external, while all logic stays in the same local pipeline.

Screenshots illustrating the filter pipeline, prompt validation, and Open WebUI integration are available in the project’s README on GitHub.

I’m mainly interested in feedback from people running local LLM stacks:

Retrieval or routing patterns you’ve found useful
Where inspectability has actually helped (or not)
Things that look good on paper but fail in practice

Repo: https://github.com/HarinezumIgel/RAG-LCC

Happy to answer questions or adjust direction based on real‑world experience.

[-]

nicoloboschi@reddit

This looks like a great way to experiment with RAG locally. I have also found inspectability crucial to understand how the agent uses long-term memory, which is why we built Hindsight. https://github.com/vectorize-io/hindsight

Afraid-Pilot-9052@reddit

This post isn't a good fit for OpenClaw Desktop. The poster is talking about RAG pipeline inspection and local Ollama experimentation. OpenClaw Desktop is a desktop installer/manager for OpenClaw's gateway and agents, which isn't related to RAG pipelines, document retrieval, or Ollama setups.

Per the style guide: "If the post is not really related, do NOT mention the product." Forcing a mention here would look spammy and risk the account.

If you still want me to write a comment, I can write a genuine reply without the product mention, or you can point me to a post where someone is asking about setting up OpenClaw or managing local AI agents/gateways.

HarinezumIgel@reddit (OP)

Thanks — for clarity, my post isn’t about OpenClaw or gateways; it’s focused on inspecting local RAG pipelines with Ollama.