hemant2747/multi-agent-framework

1# Copy to .env and adjust.2 3# --- LLM provider switch ---4# true  -> local Ollama (OLLAMA_HOST + LLM_MODEL)5# false -> cloud LLM via API key (LLM_PROVIDER/LLM_API_KEY/LLM_MODEL[/LLM_BASE_URL])6ENABLE_LOCAL_LLM=true7 8# --- Cloud LLM (used only when ENABLE_LOCAL_LLM=false) ---9LLM_PROVIDER=anthropic                 # anthropic | openai (or openai-compatible)10LLM_API_KEY=11LLM_BASE_URL=                          # optional, for openai-compatible endpoints12LLM_MAX_TOKENS=102413# cloud model id goes in LLM_MODEL below (e.g. claude-sonnet-4-6 or gpt-4o)14# cloud embeddings: EMBED_BACKEND=openai, EMBED_MODEL=text-embedding-3-small,15#                   EMBED_API_KEY (defaults to LLM_API_KEY)16 17# --- Local model (Ollama) ---18OLLAMA_HOST=http://localhost:1143419# Code-capable chat model. Smaller = faster on CPU.20#   ollama pull qwen2.5-coder:3b   (or :1.5b faster, :7b best quality)21LLM_MODEL=qwen2.5-coder:3b22# Embedding model. Pull with: ollama pull nomic-embed-text23EMBED_MODEL=nomic-embed-text24 25# Embedding backend: "ollama" (real) or "hash" (offline deterministic fallback26# so you can run / test the pipeline without Ollama installed).27EMBED_BACKEND=ollama28 29# Per-request timeouts (seconds). Raise LLM_TIMEOUT on CPU-only machines.30LLM_TIMEOUT=60031EMBED_TIMEOUT=12032 33# --- Vector store ---34EMBEDDINGS_DIR=embeddings-data35 36# --- Chunking / performance ---37CHUNK_MAX_LINES=8038CHUNK_OVERLAP_LINES=1039EMBED_BATCH_SIZE=3240 41# --- Retrieval ---42RAG_TOP_K=643 44# --- Vector index (IVF) ---45# Build an IVF (k-means) index into the JSON once a version has at least this46# many chunks; below it, search is exact brute-force. nprobe = clusters scanned47# per query (higher = better recall, slower).48IVF_MIN_RECORDS=25649IVF_NPROBE=1250 51# --- Web-search verifier (DuckDuckGo, no API key) ---52ENABLE_WEB_SEARCH=true     # concurrent verifier agent that fact-checks via web search53WEB_SEARCH_RESULTS=554WEB_SEARCH_TIMEOUT=1555 56# --- Performance ---57INGEST_CACHE=true          # reuse latest version when source unchanged58LLM_FAST_MODE=false        # single LLM call vs 5-call multi-agent flow59LLM_PARALLEL=true          # run specialists concurrently (big win for API)60SUPERVISOR_USE_LLM=false   # keyword agenda planning (no extra LLM call)61LLM_MAX_CONTEXT_CHARS=6000 # cap context sent to the LLM62LLM_KEEP_ALIVE=30m         # keep local model resident63 64# --- Optional enhancements (fall back gracefully if the lib isn't installed) ---65# Use LlamaIndex tree-sitter CodeSplitter for chunking instead of the regex one.66USE_LLAMAINDEX=false67# Run the specialist phase as a CrewAI crew (inside one LangGraph node).68USE_CREW=false69