@gnldev/rag 0.1.0

package/dist/vector-store.d.ts

@@ -0,0 +1,72 @@
+export interface VectorDoc {
+    id: string;
+    text: string;
+    metadata?: Record<string, unknown>;
+    /** 7.2: optional collection/namespace split — isolated data sets within the same store. */
+    namespace?: string;
+}
+export interface VectorItem extends VectorDoc {
+    embedding: number[];
+}
+export interface VectorMatch extends VectorDoc {
+    score: number;
+}
+/** 7.2: query narrowing/blending options (all optional → old behavior if not given). */
+export interface QueryOptions {
+    /** Only items in this namespace (matches item.namespace from upsert). All items if not given. */
+    namespace?: string;
+    /** Metadata SHALLOW equality filter: EVERY given key must match item.metadata with the SAME value. */
+    filter?: Record<string, unknown>;
+    /** Score threshold: matches whose final score falls BELOW this value are FILTERED OUT (cosine ~0..1). */
+    minScore?: number;
+    /** Hybrid search: if given, the keyword score (BM25-lite) is blended with the vector score. */
+    text?: string;
+    /** Hybrid weight w (0..1): final = (1-w)·vector + w·keyword. Default 0 → vector only (old behavior). */
+    keywordWeight?: number;
+}
+/** 7.2: delete condition — id list and/or metadata filter and/or namespace. */
+export interface DeleteWhere {
+    ids?: string[];
+    filter?: Record<string, unknown>;
+    namespace?: string;
+}
+export interface VectorStore {
+    upsert(items: VectorItem[]): Promise<void>;
+    /** 7.2: opts is backward compatible — `query(embedding, topK)` behaves identically to before if not given. */
+    query(embedding: number[], topK: number, opts?: QueryOptions): Promise<VectorMatch[]>;
+    /** 7.2 (optional): delete by id/filter/namespace, returns the number deleted. */
+    delete?(where: DeleteWhere): Promise<number>;
+}
+/** text → embedding function. Wired to the AI SDK `embed` in prod; faked in tests. */
+export type Embed = (text: string) => Promise<number[]>;
+/**
+ * Simple tokenization: lowercase, then keep runs of letters and digits.
+ *
+ * Unicode-aware on purpose. The previous character class was `[^a-z0-9çğıöşü]`, i.e. ASCII plus the
+ * six Turkish letters — so `Grüße`, `français` and `mañana` were each split at the accent and scored
+ * against fragments, silently, on any non-English corpus. @gnldev/evals' tokenizer one package over
+ * already used \p{L}/\p{N}; this is the same reading. ASCII behaviour is unchanged.
+ */
+export declare function tokenize(s: string): string[];
+/**
+ * BM25-lite keyword score (0..1): the ratio of query terms MATCHED in the document.
+ * Not full BM25 (no IDF/length normalization) — deliberately simple, "lite": a cheap,
+ * deterministic keyword signal meant to be blended with the vector score. 0 if the query is empty.
+ */
+export declare function keywordScore(query: string, doc: string): number;
+/**
+ * Metadata shallow equality: does EVERY key in filter exist in item.metadata with the same value.
+ *
+ * Exported so the other in-repo `VectorStore` implementations narrow the same way. GraphRag used to
+ * implement none of this — the second copy of a filtering rule is where the copies start to differ.
+ */
+export declare function matchesFilter(metadata: Record<string, unknown> | undefined, filter?: Record<string, unknown>): boolean;
+/** In-memory vector store (cosine similarity). The pgvector adapter implements the same interface for prod. */
+export declare class InMemoryVectorStore implements VectorStore {
+    private items;
+    upsert(items: VectorItem[]): Promise<void>;
+    query(embedding: number[], topK: number, opts?: QueryOptions): Promise<VectorMatch[]>;
+    delete(where: DeleteWhere): Promise<number>;
+}
+/** Embeds documents and writes them to the store (namespace carried over via VectorDoc). */
+export declare function indexDocuments(store: VectorStore, embed: Embed, docs: VectorDoc[]): Promise<void>;

package/dist/vector-store.js

@@ -0,0 +1,101 @@
+import { cosineSimilarity } from 'ai';
+// ── 7.2 Hybrid helpers (pure, testable) ──────────────────────────
+/**
+ * Simple tokenization: lowercase, then keep runs of letters and digits.
+ *
+ * Unicode-aware on purpose. The previous character class was `[^a-z0-9çğıöşü]`, i.e. ASCII plus the
+ * six Turkish letters — so `Grüße`, `français` and `mañana` were each split at the accent and scored
+ * against fragments, silently, on any non-English corpus. @gnldev/evals' tokenizer one package over
+ * already used \p{L}/\p{N}; this is the same reading. ASCII behaviour is unchanged.
+ */
+export function tokenize(s) {
+    return (s ?? '').toLowerCase().match(/[\p{L}\p{N}]+/gu) ?? [];
+}
+/**
+ * BM25-lite keyword score (0..1): the ratio of query terms MATCHED in the document.
+ * Not full BM25 (no IDF/length normalization) — deliberately simple, "lite": a cheap,
+ * deterministic keyword signal meant to be blended with the vector score. 0 if the query is empty.
+ */
+export function keywordScore(query, doc) {
+    const q = tokenize(query);
+    if (q.length === 0)
+        return 0;
+    const d = new Set(tokenize(doc));
+    let hit = 0;
+    for (const t of q)
+        if (d.has(t))
+            hit++;
+    return hit / q.length;
+}
+/**
+ * Metadata shallow equality: does EVERY key in filter exist in item.metadata with the same value.
+ *
+ * Exported so the other in-repo `VectorStore` implementations narrow the same way. GraphRag used to
+ * implement none of this — the second copy of a filtering rule is where the copies start to differ.
+ */
+export function matchesFilter(metadata, filter) {
+    if (!filter)
+        return true;
+    const m = metadata ?? {};
+    for (const [k, v] of Object.entries(filter)) {
+        if (m[k] !== v)
+            return false;
+    }
+    return true;
+}
+/** In-memory vector store (cosine similarity). The pgvector adapter implements the same interface for prod. */
+export class InMemoryVectorStore {
+    items = [];
+    async upsert(items) {
+        for (const it of items) {
+            const i = this.items.findIndex((x) => x.id === it.id);
+            if (i >= 0)
+                this.items[i] = it;
+            else
+                this.items.push(it);
+        }
+    }
+    async query(embedding, topK, opts) {
+        const w = opts?.keywordWeight ?? 0;
+        const hybrid = w > 0 && !!opts?.text;
+        const out = [];
+        for (const it of this.items) {
+            // 7.2: namespace + metadata narrowing (BEFORE score computation — no wasted work).
+            if (opts?.namespace !== undefined && it.namespace !== opts.namespace)
+                continue;
+            if (!matchesFilter(it.metadata, opts?.filter))
+                continue;
+            const vec = cosineSimilarity(embedding, it.embedding);
+            // 7.2: hybrid → (1-w)·vector + w·keyword; otherwise pure vector (old behavior).
+            const score = hybrid ? (1 - w) * vec + w * keywordScore(opts.text, it.text) : vec;
+            if (opts?.minScore !== undefined && score < opts.minScore)
+                continue;
+            out.push({ id: it.id, text: it.text, metadata: it.metadata, namespace: it.namespace, score });
+        }
+        return out.sort((a, b) => b.score - a.score).slice(0, topK);
+    }
+    async delete(where) {
+        const before = this.items.length;
+        const ids = where.ids ? new Set(where.ids) : undefined;
+        this.items = this.items.filter((it) => {
+            // Should it be deleted? ALL given conditions must match (ids ∧ filter ∧ namespace).
+            if (ids && !ids.has(it.id))
+                return true;
+            if (where.namespace !== undefined && it.namespace !== where.namespace)
+                return true;
+            if (where.filter && !matchesFilter(it.metadata, where.filter))
+                return true;
+            // If no condition was given (empty where), delete NOTHING (safe side).
+            if (!ids && where.namespace === undefined && !where.filter)
+                return true;
+            return false; // delete
+        });
+        return before - this.items.length;
+    }
+}
+/** Embeds documents and writes them to the store (namespace carried over via VectorDoc). */
+export async function indexDocuments(store, embed, docs) {
+    const items = await Promise.all(docs.map(async (d) => ({ ...d, embedding: await embed(d.text) })));
+    await store.upsert(items);
+}
+//# sourceMappingURL=vector-store.js.map

package/dist/vector-store.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"vector-store.js","sourceRoot":"","sources":["../src/vector-store.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,gBAAgB,EAAE,MAAM,IAAI,CAAC;AAgDtC,oEAAoE;AAEpE;;;;;;;GAOG;AACH,MAAM,UAAU,QAAQ,CAAC,CAAS;IAChC,OAAO,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC,KAAK,CAAC,iBAAiB,CAAC,IAAI,EAAE,CAAC;AAChE,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,YAAY,CAAC,KAAa,EAAE,GAAW;IACrD,MAAM,CAAC,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC;IAC1B,IAAI,CAAC,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,CAAC,CAAC;IAC7B,MAAM,CAAC,GAAG,IAAI,GAAG,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC;IACjC,IAAI,GAAG,GAAG,CAAC,CAAC;IACZ,KAAK,MAAM,CAAC,IAAI,CAAC;QAAE,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;YAAE,GAAG,EAAE,CAAC;IACvC,OAAO,GAAG,GAAG,CAAC,CAAC,MAAM,CAAC;AACxB,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,aAAa,CAAC,QAA6C,EAAE,MAAgC;IAC3G,IAAI,CAAC,MAAM;QAAE,OAAO,IAAI,CAAC;IACzB,MAAM,CAAC,GAAG,QAAQ,IAAI,EAAE,CAAC;IACzB,KAAK,MAAM,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC;QAC5C,IAAI,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC;YAAE,OAAO,KAAK,CAAC;IAC/B,CAAC;IACD,OAAO,IAAI,CAAC;AACd,CAAC;AAED,+GAA+G;AAC/G,MAAM,OAAO,mBAAmB;IACtB,KAAK,GAAiB,EAAE,CAAC;IAEjC,KAAK,CAAC,MAAM,CAAC,KAAmB;QAC9B,KAAK,MAAM,EAAE,IAAI,KAAK,EAAE,CAAC;YACvB,MAAM,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC,CAAC;YACtD,IAAI,CAAC,IAAI,CAAC;gBAAE,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,EAAE,CAAC;;gBAC1B,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QAC3B,CAAC;IACH,CAAC;IAED,KAAK,CAAC,KAAK,CAAC,SAAmB,EAAE,IAAY,EAAE,IAAmB;QAChE,MAAM,CAAC,GAAG,IAAI,EAAE,aAAa,IAAI,CAAC,CAAC;QACnC,MAAM,MAAM,GAAG,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,IAAI,CAAC;QACrC,MAAM,GAAG,GAAkB,EAAE,CAAC;QAC9B,KAAK,MAAM,EAAE,IAAI,IAAI,CAAC,KAAK,EAAE,CAAC;YAC5B,mFAAmF;YACnF,IAAI,IAAI,EAAE,SAAS,KAAK,SAAS,IAAI,EAAE,CAAC,SAAS,KAAK,IAAI,CAAC,SAAS;gBAAE,SAAS;YAC/E,IAAI,CAAC,aAAa,CAAC,EAAE,CAAC,QAAQ,EAAE,IAAI,EAAE,MAAM,CAAC;gBAAE,SAAS;YACxD,MAAM,GAAG,GAAG,gBAAgB,CAAC,SAAS,EAAE,EAAE,CAAC,SAAS,CAAC,CAAC;YACtD,gFAAgF;YAChF,MAAM,KAAK,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,GAAG,CAAC,GAAG,YAAY,CAAC,IAAK,CAAC,IAAK,EAAE,EAAE,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC;YACpF,IAAI,IAAI,EAAE,QAAQ,KAAK,SAAS,IAAI,KAAK,GAAG,IAAI,CAAC,QAAQ;gBAAE,SAAS;YACpE,GAAG,CAAC,IAAI,CAAC,EAAE,EAAE,EAAE,EAAE,CAAC,EAAE,EAAE,IAAI,EAAE,EAAE,CAAC,IAAI,EAAE,QAAQ,EAAE,EAAE,CAAC,QAAQ,EAAE,SAAS,EAAE,EAAE,CAAC,SAAS,EAAE,KAAK,EAAE,CAAC,CAAC;QAChG,CAAC;QACD,OAAO,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC;IAC9D,CAAC;IAED,KAAK,CAAC,MAAM,CAAC,KAAkB;QAC7B,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC;QACjC,MAAM,GAAG,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QACvD,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,EAAE,EAAE,EAAE;YACpC,oFAAoF;YACpF,IAAI,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,CAAC;gBAAE,OAAO,IAAI,CAAC;YACxC,IAAI,KAAK,CAAC,SAAS,KAAK,SAAS,IAAI,EAAE,CAAC,SAAS,KAAK,KAAK,CAAC,SAAS;gBAAE,OAAO,IAAI,CAAC;YACnF,IAAI,KAAK,CAAC,MAAM,IAAI,CAAC,aAAa,CAAC,EAAE,CAAC,QAAQ,EAAE,KAAK,CAAC,MAAM,CAAC;gBAAE,OAAO,IAAI,CAAC;YAC3E,uEAAuE;YACvE,IAAI,CAAC,GAAG,IAAI,KAAK,CAAC,SAAS,KAAK,SAAS,IAAI,CAAC,KAAK,CAAC,MAAM;gBAAE,OAAO,IAAI,CAAC;YACxE,OAAO,KAAK,CAAC,CAAC,SAAS;QACzB,CAAC,CAAC,CAAC;QACH,OAAO,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC;IACpC,CAAC;CACF;AAED,4FAA4F;AAC5F,MAAM,CAAC,KAAK,UAAU,cAAc,CAAC,KAAkB,EAAE,KAAY,EAAE,IAAiB;IACtF,MAAM,KAAK,GAAiB,MAAM,OAAO,CAAC,GAAG,CAC3C,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,SAAS,EAAE,MAAM,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,CAClE,CAAC;IACF,MAAM,KAAK,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;AAC5B,CAAC","sourcesContent":["import { cosineSimilarity } from 'ai';\n\nexport interface VectorDoc {\n  id: string;\n  text: string;\n  metadata?: Record<string, unknown>;\n  /** 7.2: optional collection/namespace split — isolated data sets within the same store. */\n  namespace?: string;\n}\nexport interface VectorItem extends VectorDoc {\n  embedding: number[];\n}\nexport interface VectorMatch extends VectorDoc {\n  score: number;\n}\n\n/** 7.2: query narrowing/blending options (all optional → old behavior if not given). */\nexport interface QueryOptions {\n  /** Only items in this namespace (matches item.namespace from upsert). All items if not given. */\n  namespace?: string;\n  /** Metadata SHALLOW equality filter: EVERY given key must match item.metadata with the SAME value. */\n  filter?: Record<string, unknown>;\n  /** Score threshold: matches whose final score falls BELOW this value are FILTERED OUT (cosine ~0..1). */\n  minScore?: number;\n  /** Hybrid search: if given, the keyword score (BM25-lite) is blended with the vector score. */\n  text?: string;\n  /** Hybrid weight w (0..1): final = (1-w)·vector + w·keyword. Default 0 → vector only (old behavior). */\n  keywordWeight?: number;\n}\n\n/** 7.2: delete condition — id list and/or metadata filter and/or namespace. */\nexport interface DeleteWhere {\n  ids?: string[];\n  filter?: Record<string, unknown>;\n  namespace?: string;\n}\n\nexport interface VectorStore {\n  upsert(items: VectorItem[]): Promise<void>;\n  /** 7.2: opts is backward compatible — `query(embedding, topK)` behaves identically to before if not given. */\n  query(embedding: number[], topK: number, opts?: QueryOptions): Promise<VectorMatch[]>;\n  /** 7.2 (optional): delete by id/filter/namespace, returns the number deleted. */\n  delete?(where: DeleteWhere): Promise<number>;\n}\n\n/** text → embedding function. Wired to the AI SDK `embed` in prod; faked in tests. */\nexport type Embed = (text: string) => Promise<number[]>;\n\n// ── 7.2 Hybrid helpers (pure, testable) ──────────────────────────\n\n/**\n * Simple tokenization: lowercase, then keep runs of letters and digits.\n *\n * Unicode-aware on purpose. The previous character class was `[^a-z0-9çğıöşü]`, i.e. ASCII plus the\n * six Turkish letters — so `Grüße`, `français` and `mañana` were each split at the accent and scored\n * against fragments, silently, on any non-English corpus. @gnldev/evals' tokenizer one package over\n * already used \\p{L}/\\p{N}; this is the same reading. ASCII behaviour is unchanged.\n */\nexport function tokenize(s: string): string[] {\n  return (s ?? '').toLowerCase().match(/[\\p{L}\\p{N}]+/gu) ?? [];\n}\n\n/**\n * BM25-lite keyword score (0..1): the ratio of query terms MATCHED in the document.\n * Not full BM25 (no IDF/length normalization) — deliberately simple, \"lite\": a cheap,\n * deterministic keyword signal meant to be blended with the vector score. 0 if the query is empty.\n */\nexport function keywordScore(query: string, doc: string): number {\n  const q = tokenize(query);\n  if (q.length === 0) return 0;\n  const d = new Set(tokenize(doc));\n  let hit = 0;\n  for (const t of q) if (d.has(t)) hit++;\n  return hit / q.length;\n}\n\n/**\n * Metadata shallow equality: does EVERY key in filter exist in item.metadata with the same value.\n *\n * Exported so the other in-repo `VectorStore` implementations narrow the same way. GraphRag used to\n * implement none of this — the second copy of a filtering rule is where the copies start to differ.\n */\nexport function matchesFilter(metadata: Record<string, unknown> | undefined, filter?: Record<string, unknown>): boolean {\n  if (!filter) return true;\n  const m = metadata ?? {};\n  for (const [k, v] of Object.entries(filter)) {\n    if (m[k] !== v) return false;\n  }\n  return true;\n}\n\n/** In-memory vector store (cosine similarity). The pgvector adapter implements the same interface for prod. */\nexport class InMemoryVectorStore implements VectorStore {\n  private items: VectorItem[] = [];\n\n  async upsert(items: VectorItem[]): Promise<void> {\n    for (const it of items) {\n      const i = this.items.findIndex((x) => x.id === it.id);\n      if (i >= 0) this.items[i] = it;\n      else this.items.push(it);\n    }\n  }\n\n  async query(embedding: number[], topK: number, opts?: QueryOptions): Promise<VectorMatch[]> {\n    const w = opts?.keywordWeight ?? 0;\n    const hybrid = w > 0 && !!opts?.text;\n    const out: VectorMatch[] = [];\n    for (const it of this.items) {\n      // 7.2: namespace + metadata narrowing (BEFORE score computation — no wasted work).\n      if (opts?.namespace !== undefined && it.namespace !== opts.namespace) continue;\n      if (!matchesFilter(it.metadata, opts?.filter)) continue;\n      const vec = cosineSimilarity(embedding, it.embedding);\n      // 7.2: hybrid → (1-w)·vector + w·keyword; otherwise pure vector (old behavior).\n      const score = hybrid ? (1 - w) * vec + w * keywordScore(opts!.text!, it.text) : vec;\n      if (opts?.minScore !== undefined && score < opts.minScore) continue;\n      out.push({ id: it.id, text: it.text, metadata: it.metadata, namespace: it.namespace, score });\n    }\n    return out.sort((a, b) => b.score - a.score).slice(0, topK);\n  }\n\n  async delete(where: DeleteWhere): Promise<number> {\n    const before = this.items.length;\n    const ids = where.ids ? new Set(where.ids) : undefined;\n    this.items = this.items.filter((it) => {\n      // Should it be deleted? ALL given conditions must match (ids ∧ filter ∧ namespace).\n      if (ids && !ids.has(it.id)) return true;\n      if (where.namespace !== undefined && it.namespace !== where.namespace) return true;\n      if (where.filter && !matchesFilter(it.metadata, where.filter)) return true;\n      // If no condition was given (empty where), delete NOTHING (safe side).\n      if (!ids && where.namespace === undefined && !where.filter) return true;\n      return false; // delete\n    });\n    return before - this.items.length;\n  }\n}\n\n/** Embeds documents and writes them to the store (namespace carried over via VectorDoc). */\nexport async function indexDocuments(store: VectorStore, embed: Embed, docs: VectorDoc[]): Promise<void> {\n  const items: VectorItem[] = await Promise.all(\n    docs.map(async (d) => ({ ...d, embedding: await embed(d.text) })),\n  );\n  await store.upsert(items);\n}\n"]}

package/package.json

@@ -0,0 +1,73 @@
+{
+  "name": "@gnldev/rag",
+  "version": "0.1.0",
+  "license": "Apache-2.0",
+  "engines": {
+    "node": ">=22.13.0"
+  },
+  "description": "Vector store + RAG tool for AI SDK agents. Used as a tool, it becomes replayable/exactly-once RAG with @gnldev/durable.",
+  "keywords": [
+    "ai",
+    "agent",
+    "llm",
+    "typescript",
+    "ai-sdk",
+    "durable",
+    "exactly-once",
+    "retrieval",
+    "embeddings",
+    "vector-search"
+  ],
+  "type": "module",
+  "main": "./dist/index.js",
+  "types": "./dist/index.d.ts",
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "default": "./dist/index.js"
+    },
+    "./postgres": {
+      "types": "./dist/postgres-vector-store.d.ts",
+      "default": "./dist/postgres-vector-store.js"
+    },
+    "./package.json": "./package.json"
+  },
+  "files": [
+    "dist"
+  ],
+  "peerDependencies": {
+    "ai": "^7.0.0",
+    "pg": "^8.0.0",
+    "zod": "^3.25.76 || ^4.1.8",
+    "@gnldev/durable": "^0.1.0"
+  },
+  "peerDependenciesMeta": {
+    "pg": {
+      "optional": true
+    }
+  },
+  "devDependencies": {
+    "ai": "^7.0.0",
+    "pg": "^8.0.0",
+    "zod": "^3.25.0",
+    "@gnldev/durable": "0.1.0"
+  },
+  "author": "Karaca Yılmaz (https://gnl.dev)",
+  "homepage": "https://gnl.dev",
+  "bugs": {
+    "url": "https://github.com/Karaca7/gnldev/issues"
+  },
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/Karaca7/gnldev.git",
+    "directory": "packages/rag"
+  },
+  "publishConfig": {
+    "access": "public"
+  },
+  "scripts": {
+    "build": "tsc -p tsconfig.json",
+    "typecheck": "tsc -p tsconfig.json --noEmit",
+    "test": "vitest run"
+  }
+}