Reference
Config
Environment variables and agentic-analytics.config.yaml.
Settings come from .env (gitignored; copy .env.example) and an optional agentic-analytics.config.yaml.
| Variable | Default | What |
|---|---|---|
PLANNER_MODEL, EXTRACTOR_MODEL, JUDGE_MODEL | ollama:qwen3.5:9b | ollama:<model> or openai-compatible:<endpoint>:<model> |
OLLAMA_BASE_URL | http://localhost:11434 | Ollama |
GROQ_API_KEY, NVIDIA_API_KEY, GEMINI_API_KEY, SAMBANOVA_API_KEY, MISTRAL_API_KEY, CLOUDFLARE_API_TOKEN + CLOUDFLARE_ACCOUNT_ID, OPENROUTER_API_KEY, CEREBRAS_API_KEY | — | free hosted endpoints (presets groq, nvidia-nim, google-ai-studio, sambanova, cloudflare, openrouter, cerebras) |
PLANNER_MODEL=a|b | — | the same model on several providers; the router falls back on rate limits and never switches models |
AA_MODEL_CACHE | — | response cache directory (aa eval always caches in evals/.cache) |
AA_LOG_CALLS | — | echo every model call to stderr (always logged to data/logs/model-calls.jsonl) |
EMBEDDING_MODEL | Xenova/bge-m3 | in-process embeddings (--embedder hash for the offline one) |
MODEL_TIMEOUT_MS | 120000 | per request |
SEARXNG_URL | http://localhost:8888 | the web lane |
DOCLING_URL | — | optional Docling for hard PDFs |
AA_JOBS_DIR | data/jobs | where jobs live |
API_TOKEN | — | bearer auth for the API and /mcp |
AA_ALLOWED_HOSTS | — | host names the API may be reached by besides localhost |
AA_MCP_ROOTS | — | folders analyze_files may read when MCP is called over HTTP (/mcp); none by default |
AA_ALLOW_NO_AUTH | — | 1 lets aa serve listen beyond localhost without API_TOKEN (docker-compose sets it: its port is published on 127.0.0.1 only) |
TELEGRAM_BOT_TOKEN, TELEGRAM_ALLOWED_CHATS, AA_API_URL | — | the Telegram demo (it refuses to start with no allowed chats) |
TELEGRAM_PUBLIC | — | 1 lets the Telegram bot answer any chat |
The config file’s JSON Schema: config.schema.json.