feat(ai): pgvector embedding infrastructure for semantic search
CI / backend (pull_request) Successful in 52s
CI / frontend (pull_request) Successful in 15s
Security / secrets (pull_request) Successful in 3s
Security / dependencies (pull_request) Successful in 54s

Semantic-search slice 1 (infra). Adds the pgvector plumbing the backfill worker and
hybrid search will use:
- swap the Postgres image to pgvector/pgvector:pg16 (drop-in for pg16 data)
- Pgvector + Pgvector.EntityFrameworkCore (0.2.0, EF8-compatible); UseVector() on the
  runtime + design-time contexts
- Email.Embedding vector(768) column (nomic-embed-text dims), nullable, with an HNSW
  cosine index; ignored under the InMemory test provider
- migration: CREATE EXTENSION vector + column + HNSW index

Verified against a real pgvector container: extension, HNSW, and cosine search work, and
the full EF round-trip (store a Pgvector.Vector, CosineDistance operator, nearest-first
ordering) applies all migrations and passes. No vulnerable packages. Build + all 41 tests
pass. Column stays null until Ollama generates embeddings (search falls back to lexical).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
cesnimda
2026-07-02 02:29:17 +02:00
parent d78fe601ff
commit df04568b06
10 changed files with 854 additions and 5 deletions
@@ -6,6 +6,7 @@ using Microsoft.EntityFrameworkCore.Infrastructure;
using Microsoft.EntityFrameworkCore.Storage.ValueConversion;
using Npgsql.EntityFrameworkCore.PostgreSQL.Metadata;
using NpgsqlTypes;
using Pgvector;
#nullable disable
@@ -22,6 +23,7 @@ namespace InboxIntel.Infrastructure.Migrations
.HasAnnotation("Relational:MaxIdentifierLength", 63);
NpgsqlModelBuilderExtensions.HasPostgresExtension(modelBuilder, "pg_trgm");
NpgsqlModelBuilderExtensions.HasPostgresExtension(modelBuilder, "vector");
NpgsqlModelBuilderExtensions.UseIdentityByDefaultColumns(modelBuilder);
modelBuilder.Entity("InboxIntel.Domain.Entities.AnalyticsAggregate", b =>
@@ -125,6 +127,9 @@ namespace InboxIntel.Infrastructure.Migrations
b.Property<DateTimeOffset>("CreatedAtUtc")
.HasColumnType("timestamp with time zone");
b.Property<Vector>("Embedding")
.HasColumnType("vector(768)");
b.Property<string>("GmailMessageId")
.IsRequired()
.HasMaxLength(64)
@@ -194,6 +199,11 @@ namespace InboxIntel.Infrastructure.Migrations
b.HasKey("Id");
b.HasIndex("Embedding");
NpgsqlIndexBuilderExtensions.HasMethod(b.HasIndex("Embedding"), "hnsw");
NpgsqlIndexBuilderExtensions.HasOperators(b.HasIndex("Embedding"), new[] { "vector_cosine_ops" });
b.HasIndex("SearchVector");
NpgsqlIndexBuilderExtensions.HasMethod(b.HasIndex("SearchVector"), "GIN");