local2 0.1.2

LoCAL2 — Loosely Coupled Agent Language model, second generation

LoCAL2

Loosely Coupled Agent Language model — Second Generation.

LLM-native tool calling with Gemma 4 as the orchestrator. Web search, memory recall, and feedback loops augment Gemma's native reasoning — the model decides when to use them.

Reference hardware: Mac Mini M4 Pro, 64GB unified memory. Tool calling and thinking tokens work best with sufficient VRAM/unified memory; performance on lower-spec hardware will vary.

---

Quick start

brew install ollama          # one-time prerequisite
pip install local2
local2 setup                 # pulls required models, writes initial config
local2                       # opens the web UI at http://localhost:8000

---

Prerequisites

• Python 3.11+
• Ollama — download the macOS app or brew install ollama

---

Install

pip install local2
local2 setup

local2 setup does three things:

1. Writes default config files to ~/.local2/config/
2. Pulls gemma4:e4b (generator + memory classifier)
3. Pulls nomic-embed-text (embeddings for memory and RAG library)

---

Run

local2                       # web UI, opens browser at http://localhost:8000
local2 --headless            # web server only, no browser pop
local2 --panels              # web UI + read-only Qt observer windows
local2 --desktop             # legacy PySide6 full desktop UI
local2 --model gemma4:27b    # override the generator model at startup
local2 --web-port 9000       # use a different port

---

Web search

Web search is optional. Two ways to enable it:

Option A — Brave or Tavily (no Docker required)

Get an API key from brave.com/search/api or tavily.com, then edit ~/.local2/config/web_search.yaml:

provider: brave          # or: tavily
brave_api_key: sk-...    # or tavily_api_key: tvly-...

Option B — SearXNG (self-hosted, no API key)

Requires Docker Desktop.

docker compose up -d

SearXNG runs at http://localhost:8080. This is the default provider if you don't change web_search.yaml.

You need a secret key in .env for SearXNG to start:

echo "MY_SEARX_SECRET=$(openssl rand -hex 32)" > .env

---

Academic search (optional)

search_papers uses the Semantic Scholar API. It works without a key at the free rate limit (1 req/sec). For higher limits:

echo 'export SEMANTIC_SCHOLAR_API_KEY=<your-key>' >> ~/.zshrc
source ~/.zshrc

Get a free key at semanticscholar.org/product/api.

---

Configuration

User config lives in ~/.local2/config/. Defaults are written there by local2 setup and can be edited freely — upgrades never overwrite them.

File	Controls
generator.yaml	Model, context size, temperature, tool timeout, system prompt
web_search.yaml	Search provider, API keys, max results
web_fetch.yaml	Max chars extracted, fetch timeout
critic.yaml	Critic model, grading rubric, grade timeout
memory.yaml	ChromaDB path, episodic memory collection
search_memory.yaml	Max results from memory search
semantic_scholar.yaml	Max results, request timeout
documents.yaml	Chunk size/overlap, RAG library collections
location.yaml	Optional static location override (skips live IP lookup)
bus.yaml	ZMQ proxy ports
system.yaml	Instance ID, debug flags

---

Document library (RAG)

LoCAL2 maintains a persistent local knowledge base you can query with search_library. Use the library window in the Qt UI, or the CLI:

# Ingest one or more files
PYTHONPATH=src python scripts/ingest.py path/to/file.pdf

# List all ingested sources
PYTHONPATH=src python scripts/ingest.py --list

# Delete a source by filename
PYTHONPATH=src python scripts/ingest.py --delete "file.pdf"

Supported formats: PDF, TXT, MD, PY, YAML, JSON, CSV. Files are chunked into 1500-character segments and embedded with nomic-embed-text. Re-ingesting the same file is safe — chunks are upserted by deterministic ID.

---

After a reboot

Ollama: On macOS, a stale ollama serve process can persist after reboot alongside the freshly launched Ollama.app, splitting IPv4 and IPv6 across two processes. If ollama.chat() hangs silently, check:

pgrep -fl ollama   # should show exactly one process

Kill the older PID if two appear.

SearXNG (if using Docker): Docker Desktop needs to be running before docker compose up -d.

---

Development

Clone the repo and install in editable mode:

git clone https://github.com/rkoike88/LoCAL2
cd LoCAL2
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

Run the app:

python run_local.py          # equivalent to 'local2'

Run tests:

make test                    # or: PYTHONPATH=src python -m pytest tests/ -q

Build a distributable wheel (builds frontend first):

make dist
python -m build

---

Architecture

See docs/architecture/ for design documents covering the bus topology, agent state machines, tool protocol, and configuration reference.