abmc-ollama

abmc-ollama
https://github.com/DovahkiinYuzuko/abmc-ollama

When opening a local LLM model from the command prompt, the repetitive input from

ollama list

to

ollama run <モデル>

was tedious, so I created this CLI app with the motto of "a bit more convenient." I've made installer_(OS).py for Windows, Mac, and Linux respectively, so startup should be straightforward.

You can launch it with

/abmc-ollama

■ What This Tool Can Do
[1] Run / Chat: Standard interactive mode.
[2] Add Model (Pull): Download a new model by entering its name.
[3] Remove Model (Remove): Delete unnecessary models from the list.
[4] Check Status (Status): Monitor operational status on VRAM.
[5] AI Search Chat: Automatic search mode using multi-stage RAG chains.

As you can see, since command input was tedious, I made it possible to select models by number. The AI search chat is a bonus feature—I've designed it so that even weaker models can find relevant information by having one model handle multiple roles.

■ Role Division
[1] Planner (Planning): Analyzes the user's question and creates up to 5 optimal search queries for resolution.
[2] Searcher (Collection): Uses the queries created by the Planner to actually search the web (Brave, DuckDuckGo, etc.).
(You can choose your own search engine if you input an API)
[3] Extractor (Extraction/Verification): Extracts only "facts" related to the question from the large amount of collected data. Discards outdated information by date, verifies that numerical units are correct, and carefully selects the "materials" for the answer.
[4] Writer (Writing): Based only on the "facts" selected by the Extractor, writes the final answer. Acts as an anchor to suppress AI's arbitrary assumptions (hallucinations) and produce accurate, evidence-based responses. (Not all can be prevented.)

It's not a particularly useful tool. The search function is essentially just a bonus feature. However, if Tools can be used, even a 4B model will (regardless of accuracy) investigate and compile information. Instead of breaking the session each time, I use JSON to bridge information between stages.