Checkpoint frontend UI/UX overhaul and ingestion embed benchmark work

apps/ai-service/adapters/embedding.py

@@ -1,7 +1,5 @@
 from __future__ import annotations
 
-import hashlib
-import math
 from typing import Any
 
 from rag.ports import QueryEmbeddingUnavailable
@@ -120,59 +118,3 @@ class BedrockCohereQueryEmbedder:
                 f"expected {self._dimensions}"
             )
         return values
-
-
-class SectionOnlyQueryEmbedder:
-    """Refuses to embed, confining retrieval to the route that is verified.
-
-    The section route resolves the question's attribute to a `section_key` and
-    filters on it; no vector is involved, and it measured 16/16 on the
-    human-written golden questions on 2026-08-04. Similarity measured 0.544 and
-    its provider is currently revoked.
-
-    Declining locally and immediately is better than the two alternatives it
-    replaces: a `cohere-v4` round-trip spends the boto3 retry budget before
-    failing, and `LocalHashQueryEmbedder` searches a SHA-256 vector against a
-    Cohere collection, which returns confident and meaningless hits.
-    """
-
-    def __init__(self, dimensions: int) -> None:
-        if dimensions <= 0:
-            raise ValueError("dimensions must be positive")
-        self._dimensions = dimensions
-
-    @property
-    def dimensions(self) -> int:
-        return self._dimensions
-
-    def embed_query(self, text: str) -> list[float]:  # noqa: ARG002
-        # The text is irrelevant: this embedder exists to refuse, not to embed.
-        raise QueryEmbeddingUnavailable(
-            "no query embedding provider is enabled; set EMBEDDING_PROVIDER to "
-            "use the similarity fallback"
-        )
-
-
-class LocalHashQueryEmbedder:
-    """Deterministic local plumbing probe; not a semantic retrieval model."""
-
-    def __init__(self, dimensions: int) -> None:
-        if dimensions <= 0:
-            raise ValueError("dimensions must be positive")
-        self._dimensions = dimensions
-
-    @property
-    def dimensions(self) -> int:
-        return self._dimensions
-
-    def embed_query(self, text: str) -> list[float]:
-        vector = [0.0] * self._dimensions
-        for token in text.casefold().split():
-            digest = hashlib.sha256(token.encode("utf-8")).digest()
-            index = int.from_bytes(digest[:4], "big") % self._dimensions
-            sign = 1.0 if digest[4] & 1 else -1.0
-            vector[index] += sign
-        norm = math.sqrt(sum(value * value for value in vector))
-        if norm == 0:
-            return vector
-        return [value / norm for value in vector]

apps/ai-service/adapters/postgres.py

@@ -22,6 +22,19 @@ class RetrievalTrace:
 
 
 class PostgresTraceRepository:
+    """Opens a new connection per call — no pooling (F-09: a real pool, with
+    startup-time lifecycle, is a further improvement not made here).
+
+    `connect_timeout` matters more than it looks: found live 2026-08-06 that
+    an unreachable Postgres (packets dropped, not actively refused) makes a
+    bare `psycopg.connect()` hang on the OS-level TCP timeout — tens of
+    seconds, not immediate — which defeats a caller's try/except fail-open
+    around `save()` just as effectively as no try/except at all, since the
+    exception it's waiting for never arrives in time. `routers/rag.py`
+    wraps `save()` to keep a trace outage from failing an already-computed
+    answer; this bounds how long that protection can take to kick in.
+    """
+
     def __init__(self, dsn: str) -> None:
         self._dsn = dsn
 
@@ -29,7 +42,7 @@ class PostgresTraceRepository:
         import psycopg
 
         statement = migration_path.read_text(encoding="utf-8")
-        with psycopg.connect(self._dsn) as connection:
+        with psycopg.connect(self._dsn, connect_timeout=5) as connection:
             connection.execute(statement)
 
     def save(
@@ -46,7 +59,7 @@ class PostgresTraceRepository:
         import psycopg
 
         trace_id = str(uuid.uuid4())
-        with psycopg.connect(self._dsn) as connection:
+        with psycopg.connect(self._dsn, connect_timeout=5) as connection:
             connection.execute(
                 """
                 INSERT INTO rag_retrieval_trace (
@@ -64,7 +77,7 @@ class PostgresTraceRepository:
     def get(self, trace_id: str) -> RetrievalTrace | None:
         import psycopg
 
-        with psycopg.connect(self._dsn) as connection:
+        with psycopg.connect(self._dsn, connect_timeout=5) as connection:
             row = connection.execute(
                 """
                 SELECT trace_id::text, query_text, subject_scope, query_intent,

apps/ai-service/adapters/prometheus.py

@@ -31,7 +31,10 @@ _HELP = {
     ABSTENTION: "Answers refused, by the reason retrieval gave.",
     GENERATION_REJECTED: (
         "Generations discarded before reaching the caller. `reason=\"ungrounded_number\"` "
-        "counts answers that stated a figure absent from the cited source."
+        "counts answers that stated a figure absent from the cited source; "
+        "`reason=\"uncited_claim\"` counts claims with no valid citation at all; "
+        "`reason=\"unsupported_claim\"` counts claims the entailment pass judged "
+        "not actually stated by the block they cite."
     ),
     GENERATION_SERVED: "Generations that passed grounding verification and were served.",
     ANSWER_EXTRACTIVE: "Answers served as verbatim source text.",