90 lines
4.1 KiB
Python
90 lines
4.1 KiB
Python
from __future__ import annotations
|
|
|
|
from functools import lru_cache
|
|
from pathlib import Path
|
|
|
|
from pydantic import Field
|
|
from pydantic_settings import BaseSettings, SettingsConfigDict
|
|
|
|
|
|
def _default_entities_path() -> Path:
|
|
"""2 parents up from apps/ai-service/config.py in a full repo checkout.
|
|
|
|
A container image that flattens apps/ai-service/ into its own root
|
|
(found live 2026-08-10: the deploy image does exactly this) doesn't have
|
|
that depth — ENTITIES_PATH env override is for it; this fallback just
|
|
keeps the class from crashing at import time when it's shallower.
|
|
"""
|
|
here = Path(__file__).resolve()
|
|
return (
|
|
here.parents[2] if len(here.parents) > 2 else here.parent
|
|
) / "ingestion/data/verified/drug_entities.json"
|
|
|
|
|
|
class Settings(BaseSettings):
|
|
model_config = SettingsConfigDict(env_file=".env", extra="ignore")
|
|
|
|
app_name: str = "vsf-duoc-thu-ai-service"
|
|
environment: str = "local"
|
|
qdrant_url: str = "http://localhost:6333"
|
|
qdrant_collection: str = "duocthu_v1"
|
|
qdrant_api_key: str | None = None
|
|
postgres_dsn: str = Field(
|
|
default="postgresql://duoc_thu:duoc_thu@localhost:5432/duoc_thu",
|
|
repr=False,
|
|
)
|
|
# `cohere-v4` = semantic query embedding in the corpus's own space (requires
|
|
# live Bedrock). `disabled` skips retrieval entirely. The old local-hash /
|
|
# section-only stub embedders were removed in the 2026-08-06 rebuild.
|
|
embedding_provider: str = "cohere-v4"
|
|
embedding_dimensions: int = 1024
|
|
evidence_minimum_score: float = 0.12
|
|
aws_region: str = "us-east-1"
|
|
# Generation is off unless asked for. `stub` runs the whole answer path —
|
|
# prompt, schema parsing, grounding check, fallback — with no cloud call.
|
|
# Live options: `bedrock-converse` (DeepSeek/Qwen/GLM/Nova via the Converse
|
|
# API) or `bedrock-claude` (Anthropic via the Messages path).
|
|
answer_provider: str = "disabled"
|
|
answer_model_id: str = "deepseek.v3.2"
|
|
# Optional cross-encoder rerank on the similarity fallback (needs live
|
|
# Bedrock invoke on the rerank model). Off by default; the section route
|
|
# never uses it.
|
|
rerank_enabled: bool = False
|
|
metrics_enabled: bool = True
|
|
# Optional bearer token for `GET /metrics`. Empty by default so the
|
|
# current Compose scrape and local development keep working: the endpoint
|
|
# is not reachable from the internet today (Caddy proxies only `web`, and
|
|
# ai-service publishes no host port). Set it wherever the service is
|
|
# exposed through an Ingress, which the Helm chart now allows — metrics
|
|
# carry query volumes, provider failure counts and abstain reasons.
|
|
metrics_token: str = ""
|
|
# OpenTelemetry is opt-in so the existing EC2 Compose deployment keeps
|
|
# answering when no collector is present. Docker/Kubernetes observability
|
|
# profiles enable it and point OTLP/HTTP at their local collector Service.
|
|
otel_enabled: bool = False
|
|
otel_service_name: str = "ai-service"
|
|
otel_exporter_otlp_endpoint: str = "http://localhost:4318/v1/traces"
|
|
otel_sample_ratio: float = Field(default=1.0, ge=0.0, le=1.0)
|
|
# Langfuse ingests the SAME spans over OTLP as a second exporter, so it
|
|
# adds an LLM-shaped view (per-trace answers, eval scores) without
|
|
# replacing Tempo or touching any instrumentation. All three must be set
|
|
# for the exporter to be added; keys are secrets and live only in .env /
|
|
# the cluster Secret, never in values.yaml.
|
|
langfuse_base_url: str = ""
|
|
langfuse_public_key: str = ""
|
|
langfuse_secret_key: str = ""
|
|
entities_path: Path = _default_entities_path()
|
|
# F-08: a per-turn budget across RagAgent's sequential Bedrock calls
|
|
# (understand, generate, one entailment check on the live agent path).
|
|
# Defaults sized with headroom above what a normal turn measures live
|
|
# (~8-9s before the bounded-provider change) — see rag/agent.py's
|
|
# MAX_WALL_CLOCK_MS/
|
|
# MAX_LLM_CALLS_PER_TURN for the full rationale.
|
|
max_wall_clock_ms: int = 40_000
|
|
max_llm_calls_per_turn: int = 8
|
|
|
|
|
|
@lru_cache
|
|
def get_settings() -> Settings:
|
|
return Settings()
|