Wire the guarded conversational RAG answer layer end-to-end

2026-08-05 14:33:13 +07:00
parent 834d9e51b0
commit ef08b4929e
127 changed files with 37921 additions and 169 deletions
@@ -0,0 +1,111 @@
+from __future__ import annotations
+
+from qdrant_client import QdrantClient
+
+from adapters.embedding import (
+    BedrockCohereQueryEmbedder,
+    LocalHashQueryEmbedder,
+    SectionOnlyQueryEmbedder,
+)
+from adapters.postgres import PostgresTraceRepository
+from adapters.qdrant import QdrantParentStore, QdrantRetriever
+from config import Settings
+from rag.answer import GroundedAnswerService
+from rag.artifacts import load_aliases
+from rag.conversation import DeterministicSummariser, InMemoryConversationStore
+from rag.conversational import ConversationalLoopService
+from rag.metrics import NullMetrics
+from rag.routing import CatalogDrugResolver, QueryRoutingService
+from rag.sections import SectionResolver
+from rag.service import EvidencePolicy, RetrievalService
+
+
+def _build_metrics(settings: Settings):
+    """A Prometheus exporter, or None when the package or the flag is absent.
+
+    Missing `prometheus_client` degrades to no metrics rather than to a service
+    that will not start: observability is not a precondition for answering.
+    """
+    if not settings.metrics_enabled:
+        return None
+    try:
+        from adapters.prometheus import PrometheusMetrics
+
+        return PrometheusMetrics()
+    except ImportError:
+        return None
+
+
+def _build_generator(settings: Settings):
+    if settings.answer_provider == "disabled":
+        return None
+    if settings.answer_provider == "stub":
+        from adapters.bedrock_claude import StubAnswerGenerator
+
+        return StubAnswerGenerator(
+            "Câu trả lời mẫu, không gọi nhà cung cấp nào. [1]"
+        )
+    if settings.answer_provider == "bedrock-claude":
+        from adapters.bedrock_claude import BedrockClaudeAnswerGenerator
+
+        return BedrockClaudeAnswerGenerator(
+            region=settings.aws_region, model_id=settings.answer_model_id
+        )
+    raise ValueError(
+        "Unknown ANSWER_PROVIDER. Supported values: disabled (default), "
+        "stub (local, no cloud), bedrock-claude"
+    )
+
+
+def build_runtime(settings: Settings):
+    metrics = _build_metrics(settings)
+    if settings.embedding_provider == "disabled":
+        return None, None, PostgresTraceRepository(settings.postgres_dsn), metrics
+    if settings.embedding_provider not in ("section-only", "local-smoke", "cohere-v4"):
+        raise ValueError(
+            "No production query embedder is configured. Supported values: "
+            "EMBEDDING_PROVIDER=section-only (default; section route only), "
+            "local-smoke (plumbing only) or cohere-v4"
+        )
+
+    client = QdrantClient(
+        url=settings.qdrant_url,
+        api_key=settings.qdrant_api_key,
+        timeout=30,
+    )
+    # A collection built with one model and queried with another returns hits
+    # and raises nothing; the results are just meaningless. Keep this in step
+    # with `model_id` in the collection's manifest.
+    if settings.embedding_provider == "cohere-v4":
+        embedder = BedrockCohereQueryEmbedder(
+            settings.embedding_dimensions, region=settings.aws_region
+        )
+    elif settings.embedding_provider == "local-smoke":
+        embedder = LocalHashQueryEmbedder(settings.embedding_dimensions)
+    else:
+        embedder = SectionOnlyQueryEmbedder(settings.embedding_dimensions)
+    section_resolver = SectionResolver()
+    resolver = CatalogDrugResolver(load_aliases(settings.entities_path))
+    retrieval = RetrievalService(
+        QdrantRetriever(client, settings.qdrant_collection, embedder),
+        QdrantParentStore(client, settings.qdrant_collection),
+        EvidencePolicy(minimum_score=settings.evidence_minimum_score),
+        section_resolver=section_resolver,
+    )
+    routing = QueryRoutingService(retrieval, resolver)
+    answers = GroundedAnswerService(
+        routing, generator=_build_generator(settings), metrics=metrics or NullMetrics()
+    )
+    # The conversational layer reuses the same resolvers and the safe answer
+    # engine, adding only turn understanding, follow-up inheritance and the
+    # clarify/refine loop around it. InMemory store for now; a Postgres-backed
+    # store is the persistence follow-up.
+    conversational = ConversationalLoopService(
+        answers=answers,
+        resolver=resolver,
+        section_resolver=section_resolver,
+        store=InMemoryConversationStore(),
+        summariser=DeterministicSummariser(),
+        metrics=metrics or NullMetrics(),
+    )
+    return answers, conversational, PostgresTraceRepository(settings.postgres_dsn), metrics