[app] name = "estimator" env = "dev" log_level = "INFO" [api] host = "0.0.0.0" port = 8000 request_timeout_s = 120 proxy_headers = false # true behind a reverse proxy, so client addresses are real forwarded_allow_ips = "127.0.0.1" # which proxies to trust; never "*" when the port is reachable directly [database] pool_size = 5 echo = false # dsn comes from env: EST_DATABASE__DSN [llm] repair_attempts = 1 # extra calls when JSON fails schema validation [llm.anthropic] base_url = "https://api.anthropic.com" api_version = "2023-06-01" timeout_s = 90 max_retries = 2 # api_key comes from env: EST_LLM__ANTHROPIC__API_KEY [llm.openai] base_url = "https://api.openai.com/v1" timeout_s = 90 max_retries = 2 # api_key comes from env: EST_LLM__OPENAI__API_KEY # Each pipeline step picks its own provider and model. # temperature is omitted unless set (some reasoning models reject it). [llm.steps.clarify] provider = "anthropic" model = "claude-sonnet-5-5" max_tokens = 3000 [llm.steps.decompose] provider = "anthropic" model = "claude-sonnet-5-5" max_tokens = 8000 [llm.steps.estimate] provider = "anthropic" model = "claude-sonnet-5-5" max_tokens = 8000 [llm.steps.critic] provider = "anthropic" model = "claude-sonnet-5-5" max_tokens = 6000 # Example: critic on another vendor, with the first vendor as fallback. # Check the model ID in the provider docs before use. # [llm.steps.critic] # provider = "openai" # model = "gpt-6.1-sol" # reasoning_effort = "low" # max_tokens = 6000 # [llm.steps.critic.fallback] # provider = "anthropic" # model = "claude-sonnet-5-5" # max_tokens = 6000 [agent] prompts_dir = "prompts" max_requirements_chars = 20000 max_questions = 7 max_tasks = 60 max_task_hours = 400 max_critic_new_tasks = 10 critic_max_change_ratio = 3.0 [rag] # Retrieval of similar tasks from history. Load a corpus first: python -m estimator rag import FILE enabled = false provider = "openai" # openai (any OpenAI-compatible /embeddings endpoint) | hash (offline, weak) model = "text-embedding-3-small" # the vector size is fixed at 1536 by the database schema base_url = "https://api.openai.com/v1" timeout_s = 60 max_retries = 2 batch_size = 64 top_k = 3 # references shown per task min_similarity = 0.55 # api_key comes from env: EST_RAG__API_KEY (falls back to EST_LLM__OPENAI__API_KEY) [pricing] default_region = "ru" default_currency = "RUB" default_engagement = "fixed_price" # fixed_price | time_and_materials, chosen per estimate in the UI/API pm_pct = 0.15 markup_pct = 0.20 # markup on cost; the implied margin is markup / (1 + markup) reserve_floor_pct = 0.10 # minimum reserve; the PERT reserve is used when it is larger reserve_z = 1.04 # PERT reserve = reserve_z x sigma of the estimate, not a calibrated quantile estimate_unit_hours_per_day = 8 [pricing.vat] # VAT is shown on top of prices that are calculated without it. Defaults by the client's region: # Russian client: 22% (the general rate since 2026, for a seller that charges VAT); EU client: none, # services to a foreign client are outside Russian VAT. Override per estimate in the UI or API. default_rates = { ru = 0.22, eu = 0.0 } choices = [0.22, 0.10, 0.07, 0.05] # other rates offered besides "no VAT" [pricing.min_rounding] step = 1000 [market] active_version = "2026Q4" check_tolerance = 0.25 [calibration] enabled = true min_samples = 5 default_k = 1.2 [fx] provider = "cbr" cache_ttl_s = 3600 [mcp] enabled = false transport = "stdio" [web] enabled = false api_base_url = "" # empty = call the API inside this process session_ttl_s = 28800 page_size = 20 poll_interval_s = 2 workspace = "shared" # shared | per_user timezone = "UTC" cookie_secure = false # true behind HTTPS login_max_attempts = 5 login_window_s = 300 # secrets come from env: EST_WEB__SESSION_SECRET, EST_WEB__USERS, EST_WEB__INTERNAL_API_KEY [audit] store_prompts = true [jobs] max_concurrent = 4 timeout_s = 300 shutdown_grace_s = 10