Checkpoint frontend UI/UX overhaul and ingestion embed benchmark work

This commit is contained in:
2026-08-06 17:21:21 +07:00
parent 1e8cbdb586
commit a4b8e1c4db
78 changed files with 6761 additions and 654 deletions
+90 -34
View File
@@ -2,22 +2,51 @@ from __future__ import annotations
from qdrant_client import QdrantClient
from adapters.embedding import (
BedrockCohereQueryEmbedder,
LocalHashQueryEmbedder,
SectionOnlyQueryEmbedder,
)
from adapters.embedding import BedrockCohereQueryEmbedder
from adapters.postgres import PostgresTraceRepository
from adapters.qdrant import QdrantParentStore, QdrantRetriever
from config import Settings
from rag.agent import RagAgent
from rag.answer import GroundedAnswerService
from rag.artifacts import load_aliases
from rag.conversation import DeterministicSummariser, InMemoryConversationStore
from rag.conversational import ConversationalLoopService
from rag.manifest import MANIFEST_POINT_ID, check_manifest, manifest_collection
from rag.metrics import NullMetrics
from rag.routing import CatalogDrugResolver, QueryRoutingService
from rag.sections import SectionResolver
from rag.service import EvidencePolicy, RetrievalService
from rag.understanding import LlmQueryUnderstander
# How many aliases to show per candidate drug (F-04 bounds *which* drugs are
# shown at all, per turn; this bounds how many names each shown one gets).
_CATALOG_ALIASES_PER_DRUG = 3
def _catalog_names(aliases: dict[str, set[str]]) -> dict[str, str]:
"""The name(s) shown to the LLM for each drug_id.
Found live 2026-08-06: picking the first 3 aliases *alphabetically* could
drop the drug's own canonical/INN name entirely — paracetamol has 191
aliases (mostly trade names), and the alphabetically-first 3 were
"0Frezefev, ABAB, Ace kid 80", none recognizable. Mid-conversation, once
the current turn no longer restates the drug name in raw text, the model
has to re-derive it from history + this catalog line alone — with no
recognizable name shown, it read the earlier "paracetamol" mention as an
unknown drug and answered "not found in the formulary" for a drug that
plainly is. Fixed by always showing the drug_id's own name form first
(guaranteed present, always recognizable — it's the exact string a user
who names a drug is most likely to type), then filling remaining slots
with short, ALL-CAPS-preferring aliases (the book's own heading
convention, so usually the generic name, not a dosage-suffixed brand
like "Ace kid 80").
"""
result: dict[str, str] = {}
for drug_id, names in aliases.items():
canonical = drug_id.replace("_", " ")
ranked = sorted(names, key=lambda n: (not n.isupper(), len(n), n))
extra = [n for n in ranked if n.strip().casefold() != canonical]
shown = [canonical, *extra][:_CATALOG_ALIASES_PER_DRUG]
result[drug_id] = ", ".join(dict.fromkeys(shown))
return result
def _build_metrics(settings: Settings):
@@ -73,15 +102,37 @@ def _build_reranker(settings: Settings):
return BedrockCohereReranker(region=settings.aws_region)
def _verify_corpus_manifest(client, collection: str, embedder, settings: Settings) -> None:
"""Read the sidecar manifest point the ingestion loader writes
(`ingestion/ingestion/load/manifest.py`) and refuse to start on a
mismatch. Raises `rag.manifest.ManifestMismatch` (a `RuntimeError`),
which crashes startup — a deliberate refusal, not an oversight: this
runs at process start (`main.py` calls `build_runtime` at import time),
so a mismatch here means the process never comes up and never serves a
query from a corpus it wasn't verified against.
"""
sidecar = manifest_collection(collection)
payload = None
# This qdrant-client version has no `collection_exists`, and
# `get_collection` (singular) is a known parse-bug risk in this
# environment — list collections and check membership instead.
existing = {col.name for col in client.get_collections().collections}
if sidecar in existing:
points = client.retrieve(sidecar, [MANIFEST_POINT_ID], with_payload=True)
if points:
payload = points[0].payload
check_manifest(payload, collection, embedder.model_id, settings.embedding_dimensions)
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"):
if settings.embedding_provider != "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"
"EMBEDDING_PROVIDER=cohere-v4 (semantic query embedding) or "
"disabled. The old local/section-only stubs were removed."
)
client = QdrantClient(
@@ -89,19 +140,21 @@ def build_runtime(settings: Settings):
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)
embedder = BedrockCohereQueryEmbedder(
settings.embedding_dimensions, region=settings.aws_region
)
# F-05: a collection built with one model and queried with another
# returns hits and raises nothing — the results are just meaningless,
# with no error to notice. Refuse to start rather than search with
# vectors this collection was not built from.
_verify_corpus_manifest(client, settings.qdrant_collection, embedder, settings)
section_resolver = SectionResolver()
resolver = CatalogDrugResolver(load_aliases(settings.entities_path))
aliases = load_aliases(settings.entities_path)
# Kept only for deterministic input-time autocomplete (`RagAgent.complete`)
# and to satisfy `GroundedAnswerService`'s constructor — its fuzzy
# `.resolve()` is no longer on the live query path; `RagAgent` resolves
# drug identity through `LlmQueryUnderstander` against the same catalog.
resolver = CatalogDrugResolver(aliases)
retrieval = RetrievalService(
QdrantRetriever(client, settings.qdrant_collection, embedder),
QdrantParentStore(client, settings.qdrant_collection),
@@ -110,19 +163,22 @@ def build_runtime(settings: Settings):
reranker=_build_reranker(settings),
)
routing = QueryRoutingService(retrieval, resolver)
generator = _build_generator(settings)
answers = GroundedAnswerService(
routing, generator=_build_generator(settings), metrics=metrics or NullMetrics()
routing, generator=generator, 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(
trace_writer = PostgresTraceRepository(settings.postgres_dsn)
if generator is None:
# The new front end understands a turn with the same LLM call that
# answers it — with no generator configured there is no query
# understanding either, so there is no conversational/agent
# capability to offer. Answer-only (retrieval-verified, no
# generation) still works through `answers` directly.
return answers, None, trace_writer, metrics
agent = RagAgent(
understander=LlmQueryUnderstander(generator, _catalog_names(aliases), resolver),
retrieval=retrieval,
answers=answers,
resolver=resolver,
section_resolver=section_resolver,
store=InMemoryConversationStore(),
summariser=DeterministicSummariser(),
metrics=metrics or NullMetrics(),
autocomplete=resolver,
)
return answers, conversational, PostgresTraceRepository(settings.postgres_dsn), metrics
return answers, agent, trace_writer, metrics