Imported from adhitiad/fr (AGENTS.md). Install upstream with npx skills add adhitiad/fr. Copyright stays with the author.

AGENTS.md

Project

Multi-agent AI orchestrator (Indonesian). Entry point: main.py. Three sub-agents (coder, admin, casual) route via keyword matching. Uses LangChain, NVIDIA NIM LLM, Redis memory, FAISS RAG.

Run

• Install deps: pip install -r requirements.txt (venv at venv/)
• Copy .env.example → .env, set NVIDIA_API_KEY (must not be the placeholder string)
• Run: python main.py
• Redis must be running locally (redis://localhost:6379/0) — memory layer depends on it

Architecture

• main.py — orchestrator. Keyword router at line 52-60. Every route_request() call holds api_lock and sleeps 3s (deliberate rate limit).
• agents/ — coder_agent.py, admin_agent.py, casual_agent.py. Each builds its own LLM chain with tool binding + Redis memory + RAG context.
• agents/agent.py, agents/env.py — placeholder RL framework (not yet implemented).
• agents/llm_config.py — centralized LLM setup. Uses ChatNVIDIA (NVIDIA NIM). Model defaults to moonshotai/kimi-k3, overridden by LLM_MODEL in .env.
• config/rules.py — source of truth for agent roles, rules, and which tools (skills) each agent gets. MCP registry reads this.
• mcp/registry.py — load_mcp_context(agent_type) maps agent → tools from AVAILABLE_PLUGINS in plugins/core_tools.py.
• mcp/context_loader.py — loads mcp/mcp.json, rules/*.md, skills/*.md into a combined context blob injected into system prompts.
• plugins/core_tools.py — real tool definitions: tulis_kode (write file), baca_file (read file), cari_web (DuckDuckGo search).
• tools/agent_tools.py — legacy duplicate of plugins/core_tools.py. Do not import from here; use plugins/.
• tools/rag_engine.py — FAISS + HuggingFace embeddings (all-MiniLM-L6-v2). Requires faiss_index/ to exist. Build it first:

  from tools.rag_engine import ingest_documents
  ingest_documents()

• memory/memory.py — Redis-backed chat history per session_id.
• core/logger.py — standard logging setup.

Critical Quirks

• Redis monkey-patch: main.py patches redis.from_url at import time to force protocol=2 (line 8-18). If you work with Redis directly, import main first or replicate the patch.
• FAISS index required: get_relevant_context() raises RuntimeError if faiss_index/ is missing. Run ingest_documents() before any agent call that uses RAG.
• .env validation: llm_config.py raises RuntimeError if NVIDIA_API_KEY is missing or equals the placeholder your_gemini_api_key_here.
• Rate limit: route_request() sleeps 3s per call under a lock. Expect slow sequential runs.
• Duplicate tools: tools/agent_tools.py duplicates plugins/core_tools.py. Agents import from plugins/. Ignore tools/agent_tools.py.
• No test framework: Tests are ad-hoc scripts (test_concurrent.py, test_redis_patch.py). Run with python test_concurrent.py. No pytest, no lint, no typecheck config exists.

Routing Keywords (case-insensitive)

Data

• data/*.txt — RAG source documents (e.g. kebijakan_umk.txt, ai_trading_guidelines.txt). Re-ingest after editing.
• skills/coding.md, rules/policy.md — markdown injected into agent context. Currently placeholder content.
• mcp/mcp.json — empty {}. MCP server config lives here when populated.