fix: remove pgvector dependency from AI knowledge retrieval

.trellis/spec/backend/ai-sdk-integration.md

@@ -350,7 +350,7 @@ GOOGLE_GENERATIVE_AI_API_KEY=...
 ANTHROPIC_API_KEY=sk-ant-...
 ```
 
-## Scenario: Elder AI Analysis MVP With LangChain And pgvector
+## Scenario: Elder AI Analysis MVP With LangChain And PostgreSQL JSONB Embeddings
 
 ### 1. Scope / Trigger
 
@@ -369,7 +369,7 @@ ANTHROPIC_API_KEY=sk-ant-...
   - `DELETE /api/ai/knowledge/[id]`
 - DB:
   - `ai_knowledge_entries`
-  - `ai_knowledge_chunks` with `vector(1536)`
+  - `ai_knowledge_chunks` with `jsonb` `embedding` arrays
   - `elder_ai_analyses`
 - Runtime env:
   - `AI_API_KEY` required for generation and embedding.
@@ -394,6 +394,7 @@ ANTHROPIC_API_KEY=sk-ant-...
   - `errorCategory` / `errorReason`: sanitized only for failed rows
 - Failed rows must not store prompts, full resident context snapshots, API keys, or raw provider errors.
 - Knowledge retrieval may return enabled platform knowledge and enabled active-organization knowledge only.
+- Knowledge retrieval first filters enabled platform / active-organization chunks in SQL, then computes cosine similarity in `modules/ai/server/knowledge.ts`; this keeps the MVP deployable on plain PostgreSQL without the `vector` extension.
 - Completed elder analysis `resultJson.citations` must be derived from citation IDs actually referenced by `keyFindings[].citationIds` and `recommendations[].citationIds`; do not append unused available citations.
 
 ### 4. Validation & Error Matrix
@@ -420,7 +421,7 @@ ANTHROPIC_API_KEY=sk-ant-...
 ### 6. Tests Required
 
 - Permission seeding assertions for `ai:*` and `knowledge:*` default role grants.
-- Knowledge retrieval tests for platform shared, own organization, other organization, and disabled entry filtering.
+- Knowledge retrieval tests for platform shared, own organization, other organization, disabled entry filtering, and cosine-score ordering from JSONB embeddings.
 - Analysis tests for missing config, schema validation failure, failed-history sanitization, and data-scope redaction.
 - Analysis tests for unknown citation rejection and referenced-only citation persistence.
 - Route tests for auth/permission failures and standard response shape.

.trellis/tasks/07-05-07-06-harden-ai-analysis-path/design.md

@@ -63,8 +63,8 @@ Tests should mock database/provider boundaries and assert observable contracts,
 
 ## Compatibility
 
-No database migration is required for this hardening pass.
 No new dependency is required.
+Deployment hardening additionally changes the still-unapplied `0007` AI migration from pgvector to JSONB embeddings, so production PostgreSQL 18 Alpine can run the migration without a `vector` extension.
 Existing persisted `resultJson.citations` remain readable; new analyses will have stricter citation lists.
 
 ## Rollback

.trellis/tasks/07-05-07-06-harden-ai-analysis-path/implement.md

@@ -9,7 +9,8 @@
 5. Add AI route tests for permission failures, invalid input, service failure propagation, success responses, and audit calls.
 6. Improve knowledge UI and analysis dialog failure copy.
 7. Clarify `ai:manage` permission description.
-8. Run targeted AI tests, then project verification.
+8. Align AI knowledge embedding storage to JSONB for production PostgreSQL without pgvector.
+9. Run targeted AI tests, then project verification.
 
 ## Validation Commands
 

.trellis/tasks/07-05-07-06-harden-ai-analysis-path/prd.md

@@ -6,13 +6,13 @@ Harden the existing elder AI analysis and knowledge retrieval MVP so the critica
 
 ## Background
 
-The current MVP already implements LangChain-backed elder AI analysis, pgvector-backed knowledge retrieval, permission-scoped resident context, persisted analysis history, and knowledge management UI.
+The current MVP now implements LangChain-backed elder AI analysis, PostgreSQL JSONB-backed knowledge retrieval, permission-scoped resident context, persisted analysis history, and knowledge management UI.
 
 Current evidence from repository inspection:
 
 - AI service files live under `modules/ai/`.
 - API routes live under `app/api/ai/`.
-- AI tables and vector extension are in `modules/core/server/schema.ts` and `drizzle/0007_purple_puff_adder.sql`.
+- AI tables and JSONB embedding storage are in `modules/core/server/schema.ts` and `drizzle/0007_purple_puff_adder.sql`; the production target does not require PostgreSQL `vector` extension support.
 - Existing project verification passes: `pnpm lint`, `pnpm type-check`, and `pnpm test`.
 - No AI-specific test file currently covers schema validation, permission scoping, history redaction, knowledge organization isolation, disabled knowledge exclusion, or provider failure behavior.
 - `normalizeCitations` currently appends allowed citations that the model did not necessarily cite, weakening the evidence chain.

drizzle/0007_purple_puff_adder.sql

@@ -1,4 +1,3 @@
-CREATE EXTENSION IF NOT EXISTS vector;--> statement-breakpoint
 CREATE TYPE "public"."ai_knowledge_scope" AS ENUM('platform', 'organization');--> statement-breakpoint
 CREATE TYPE "public"."ai_knowledge_status" AS ENUM('enabled', 'disabled');--> statement-breakpoint
 CREATE TYPE "public"."elder_ai_analysis_status" AS ENUM('completed', 'failed');--> statement-breakpoint
@@ -9,7 +8,7 @@ CREATE TABLE "ai_knowledge_chunks" (
 	"scope" "ai_knowledge_scope" NOT NULL,
 	"chunk_index" integer NOT NULL,
 	"content" text NOT NULL,
-	"embedding" vector(1536) NOT NULL,
+	"embedding" jsonb DEFAULT '[]'::jsonb NOT NULL,
 	"source_title" text NOT NULL,
 	"source_category" text DEFAULT '' NOT NULL,
 	"created_at" timestamp with time zone DEFAULT now() NOT NULL
@@ -55,7 +54,6 @@ ALTER TABLE "elder_ai_analyses" ADD CONSTRAINT "elder_ai_analyses_elder_id_elder
 ALTER TABLE "elder_ai_analyses" ADD CONSTRAINT "elder_ai_analyses_actor_account_id_accounts_id_fk" FOREIGN KEY ("actor_account_id") REFERENCES "public"."accounts"("id") ON DELETE set null ON UPDATE no action;--> statement-breakpoint
 CREATE INDEX "ai_knowledge_chunks_entry_lookup" ON "ai_knowledge_chunks" USING btree ("entry_id");--> statement-breakpoint
 CREATE INDEX "ai_knowledge_chunks_scope_lookup" ON "ai_knowledge_chunks" USING btree ("scope","organization_id");--> statement-breakpoint
-CREATE INDEX "ai_knowledge_chunks_embedding_cosine_idx" ON "ai_knowledge_chunks" USING hnsw ("embedding" vector_cosine_ops);--> statement-breakpoint
 CREATE INDEX "ai_knowledge_entries_scope_lookup" ON "ai_knowledge_entries" USING btree ("scope","status");--> statement-breakpoint
 CREATE INDEX "ai_knowledge_entries_organization_lookup" ON "ai_knowledge_entries" USING btree ("organization_id","status");--> statement-breakpoint
 CREATE INDEX "elder_ai_analyses_elder_lookup" ON "elder_ai_analyses" USING btree ("organization_id","elder_id","created_at");--> statement-breakpoint

drizzle/meta/0007_snapshot.json

@@ -277,9 +277,10 @@
         },
         "embedding": {
           "name": "embedding",
-          "type": "vector(1536)",
+          "type": "jsonb",
           "primaryKey": false,
-          "notNull": true
+          "notNull": true,
+          "default": "'[]'::jsonb"
         },
         "source_title": {
           "name": "source_title",

modules/ai/server/knowledge.test.ts

@@ -130,7 +130,7 @@ type RetrievedChunkRow = {
   title: string;
   category: string;
   content: string;
-  distance: number;
+  embedding: number[];
 };
 
 type KnowledgeDatabaseFake = {
@@ -325,7 +325,7 @@ class SelectBuilder implements PromiseLike<SelectRow[]> {
           title: chunk.sourceTitle,
           category: chunk.sourceCategory,
           content: chunk.content,
-          distance: 0.1 + rows.length * 0.1,
+          embedding: chunk.embedding,
         });
       }
     }
@@ -407,6 +407,31 @@ describe("AI knowledge service", () => {
     expect(embeddingMocks.embedDocuments).toHaveBeenCalledWith(["night fall risk"]);
   });
 
+  it("orders retrieved chunks by cosine similarity after scope filtering", async () => {
+    const weakEntry = createEntry({ id: "entry-weak", title: "Weak guidance" });
+    const strongEntry = createEntry({ id: "entry-strong", title: "Strong guidance" });
+    useDatabase(createDatabaseFake({
+      entries: [weakEntry, strongEntry],
+      chunks: [
+        createChunk(weakEntry, { embedding: [0, 1, 0] }),
+        createChunk(strongEntry, { embedding: [1, 0, 0] }),
+      ],
+    }));
+    embeddingMocks.embedDocuments.mockResolvedValueOnce([[1, 0, 0]]);
+
+    const result = await retrieveKnowledge("org-1", "night fall risk", { limit: 2 });
+
+    expect(result).toEqual({
+      success: true,
+      data: {
+        results: [
+          expect.objectContaining({ entryId: "entry-strong", score: 1 }),
+          expect.objectContaining({ entryId: "entry-weak", score: 0 }),
+        ],
+      },
+    });
+  });
+
   it("requires platform management permission before mutating platform knowledge", async () => {
     const existingPlatformEntry = createEntry({ id: "entry-platform", scope: "platform", organizationId: null });
     useDatabase(createDatabaseFake({ entries: [existingPlatformEntry] }));

modules/ai/server/knowledge.ts

@@ -1,7 +1,7 @@
 import "server-only";
 
 import { OpenAIEmbeddings } from "@langchain/openai";
-import { and, desc, eq, isNull, or, sql } from "drizzle-orm";
+import { and, desc, eq, isNull, or } from "drizzle-orm";
 
 import { getAiRuntimeConfig } from "@/modules/ai/server/config";
 import type {
@@ -316,6 +316,29 @@ export async function deleteKnowledgeEntry(context: AuthContext, id: string): Pr
   return { success: true, data: { entry: toKnowledgeEntry(row) } };
 }
 
+function calculateCosineSimilarity(left: readonly number[], right: readonly number[]): number {
+  if (left.length === 0 || right.length === 0 || left.length !== right.length) {
+    return 0;
+  }
+
+  let dotProduct = 0;
+  let leftMagnitude = 0;
+  let rightMagnitude = 0;
+  for (let index = 0; index < left.length; index += 1) {
+    const leftValue = left[index] ?? 0;
+    const rightValue = right[index] ?? 0;
+    dotProduct += leftValue * rightValue;
+    leftMagnitude += leftValue * leftValue;
+    rightMagnitude += rightValue * rightValue;
+  }
+
+  if (leftMagnitude === 0 || rightMagnitude === 0) {
+    return 0;
+  }
+
+  return dotProduct / (Math.sqrt(leftMagnitude) * Math.sqrt(rightMagnitude));
+}
+
 export async function retrieveKnowledge(
   organizationId: string,
   query: string,
@@ -333,7 +356,6 @@ export async function retrieveKnowledge(
   }
 
   const database = getDatabase();
-  const distance = sql<number>`${aiKnowledgeChunks.embedding} <=> ${JSON.stringify(embedding)}::vector`;
   const rows = await database
     .select({
       entryId: aiKnowledgeChunks.entryId,
@@ -341,25 +363,26 @@ export async function retrieveKnowledge(
       title: aiKnowledgeChunks.sourceTitle,
       category: aiKnowledgeChunks.sourceCategory,
       content: aiKnowledgeChunks.content,
-      distance,
+      embedding: aiKnowledgeChunks.embedding,
     })
     .from(aiKnowledgeChunks)
     .innerJoin(aiKnowledgeEntries, eq(aiKnowledgeEntries.id, aiKnowledgeChunks.entryId))
-    .where(getEnabledChunkWhere(organizationId))
-    .orderBy(distance)
-    .limit(options?.limit ?? DEFAULT_RETRIEVAL_LIMIT);
+    .where(getEnabledChunkWhere(organizationId));
 
   return {
     success: true,
     data: {
-      results: rows.map((row) => ({
+      results: rows
+        .map((row) => ({
           entryId: row.entryId,
           chunkId: row.chunkId,
           title: row.title,
           category: row.category,
           content: row.content,
-        score: 1 - row.distance,
-      })),
+          score: calculateCosineSimilarity(row.embedding, embedding),
+        }))
+        .sort((left, right) => right.score - left.score)
+        .slice(0, options?.limit ?? DEFAULT_RETRIEVAL_LIMIT),
     },
   };
 }

modules/core/server/schema.ts

@@ -11,7 +11,6 @@ import {
   timestamp,
   uniqueIndex,
   uuid,
-  vector,
 } from "drizzle-orm/pg-core";
 
 export const accountStatusEnum = pgEnum("account_status", ["active", "disabled", "pending"]);
@@ -555,7 +554,7 @@ export const aiKnowledgeChunks = pgTable("ai_knowledge_chunks", {
   scope: aiKnowledgeScopeEnum("scope").notNull(),
   chunkIndex: integer("chunk_index").notNull(),
   content: text("content").notNull(),
-  embedding: vector("embedding", { dimensions: AI_KNOWLEDGE_EMBEDDING_DIMENSIONS }).notNull(),
+  embedding: jsonb("embedding").$type<number[]>().notNull().default(sql`'[]'::jsonb`),
   sourceTitle: text("source_title").notNull(),
   sourceCategory: text("source_category").notNull().default(""),
   createdAt: timestamp("created_at", { withTimezone: true }).notNull().defaultNow(),