agentfootprint 2.0.1 → 2.1.0

package/dist/types/lib/rag/defineRAG.d.ts

@@ -0,0 +1,141 @@
+/**
+ * defineRAG — sugar factory for retrieval-augmented generation.
+ *
+ * RAG is a context-engineering flavor: embed the user's question,
+ * retrieve top-K semantically similar chunks from a vector store,
+ * inject those chunks into the messages slot of the next LLM call.
+ * It's the same plumbing as `defineMemory({ type: SEMANTIC,
+ * strategy: TOP_K })` — the rename is for intent + ergonomics.
+ *
+ *   defineMemory ─┬─► EPISODIC   (raw conversation)
+ *                 ├─► SEMANTIC   (extracted facts / RAG chunks)
+ *                 ├─► NARRATIVE  (beats / summaries)
+ *                 └─► CAUSAL     (footprintjs decision snapshots)
+ *
+ *   defineRAG    ─►  SEMANTIC + TOP_K with RAG-specific defaults
+ *                    (asRole='user', threshold=0.7, no LLM-extract)
+ *
+ * Pattern: Composition over duplication — defineRAG returns a
+ *          MemoryDefinition produced by defineMemory. No new engine
+ *          code, no new slot subflow, no new event type.
+ *
+ * Role:    Layer-3 context-engineering primitive. Lives next to
+ *          defineSkill / defineSteering / defineInstruction / defineFact
+ *          but resolves to a memory subflow rather than an Injection
+ *          (RAG content is computed at runtime via async retrieval —
+ *          can't fit the synchronous Injection.inject shape).
+ *
+ * Emits:   Indirectly — the underlying memory pipeline emits
+ *          `agentfootprint.context.injected` when retrieved chunks
+ *          land in the messages slot.
+ *
+ * 7-panel review (2026-04-29):
+ * - LLM Systems    ✅  injects as 'user' role by default — RAG chunks
+ *                      land where the LLM treats them as authoritative
+ *                      retrieved context, not behavior rules
+ * - Architect      ✅  composition over defineMemory; zero new engine
+ *                      code; multi-RAG layering works via per-id keys
+ * - API Designer   ✅  one factory, mirrors defineMemory shape; consumer
+ *                      ergonomics: `agent.rag(defineRAG({...}))`
+ * - Performance    ✅  embedding cost is one call per turn (TURN_START
+ *                      timing, default); strict threshold prevents
+ *                      injecting low-confidence noise
+ * - Privacy        ✅  multi-tenant via MemoryIdentity tuple; doc
+ *                      content never crosses tenant boundaries
+ * - ML / IR        ✅  embedder version pinned via `embedderId`; cosine
+ *                      score semantics inherited from MemoryStore
+ * - SoftEng        ✅  thin file (this one); existing memory tests
+ *                      cover the underlying pipeline
+ *
+ * @see ./indexDocuments.ts  for the seeding helper
+ * @see ../../memory/define.ts  for the underlying factory
+ *
+ * @example  Basic usage
+ * ```ts
+ * import {
+ *   Agent, defineRAG, indexDocuments, InMemoryStore, mockEmbedder,
+ * } from 'agentfootprint';
+ *
+ * const embedder = mockEmbedder();
+ * const store = new InMemoryStore();
+ *
+ * // Seed the store once at startup
+ * await indexDocuments(store, embedder, [
+ *   { id: 'doc1', content: 'Refunds are processed within 3 business days.' },
+ *   { id: 'doc2', content: 'Pro plan costs $20/month.' },
+ * ]);
+ *
+ * const docs = defineRAG({
+ *   id: 'product-docs',
+ *   description: 'Retrieve product documentation chunks',
+ *   store,
+ *   embedder,
+ *   topK: 3,
+ *   threshold: 0.6,
+ * });
+ *
+ * const agent = Agent.create({ provider }).rag(docs).build();
+ * ```
+ */
+import type { ContextRole } from '../../events/types.js';
+import type { Embedder } from '../../memory/embedding/index.js';
+import type { MemoryStore } from '../../memory/store/index.js';
+import type { MemoryDefinition } from '../../memory/define.types.js';
+export interface DefineRAGOptions {
+    /** Stable id. Becomes the scope-key suffix and the Lens label. */
+    readonly id: string;
+    /**
+     * Human-readable description. Surfaces in narrative + Lens hover.
+     * Recommend describing the *corpus* (e.g., "Product documentation
+     * chunks indexed weekly from docs.example.com").
+     */
+    readonly description?: string;
+    /**
+     * Vector-capable store containing the indexed corpus. Must implement
+     * `search()`. Use `indexDocuments(store, embedder, docs)` at startup
+     * to populate it. Ships with `InMemoryStore` for dev/tests; swap to
+     * `pgvector` / Pinecone / Qdrant adapters in production.
+     */
+    readonly store: MemoryStore;
+    /**
+     * Embedder used for the read-side query. Pass the SAME embedder
+     * instance (or one with the same `name`) that was used for indexing
+     * — cross-model similarity scores are not comparable.
+     */
+    readonly embedder: Embedder;
+    /**
+     * Stable id of the embedder. Stored on entries during indexing
+     * (via `indexDocuments`) and filtered at search time so a later
+     * embedder swap doesn't pollute results.
+     */
+    readonly embedderId?: string;
+    /**
+     * Top-K chunks to retrieve per turn. Default 3 (balanced —
+     * defends against lost-in-the-middle while giving multiple
+     * perspectives). Increase for richer context, decrease for cost.
+     */
+    readonly topK?: number;
+    /**
+     * Minimum cosine similarity to inject. **Strict** — when no chunk
+     * meets the threshold, NO injection happens (no fallback that would
+     * pollute the prompt with weak matches). Default 0.7.
+     */
+    readonly threshold?: number;
+    /**
+     * Role to use when injecting retrieved chunks into the messages
+     * slot. Default `'user'` — RAG chunks are most often treated as
+     * "context the user provided" by LLMs. Use `'system'` for
+     * authoritative reference docs that should outweigh user instruction.
+     */
+    readonly asRole?: ContextRole;
+}
+/**
+ * Build a RAG context-engineering definition. The returned
+ * `MemoryDefinition` is registered on the Agent via `.rag(definition)`
+ * (or, equivalently, `.memory(definition)` — same plumbing).
+ *
+ * @throws when `store` does not implement `search()`. RAG requires a
+ *         vector-capable adapter.
+ */
+export declare function defineRAG(opts: DefineRAGOptions): MemoryDefinition;
+//# sourceMappingURL=defineRAG.d.ts.map

package/dist/types/lib/rag/defineRAG.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"defineRAG.d.ts","sourceRoot":"","sources":["../../../../src/lib/rag/defineRAG.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8EG;AAEH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,uBAAuB,CAAC;AACzD,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,iCAAiC,CAAC;AAChE,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,6BAA6B,CAAC;AAC/D,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,8BAA8B,CAAC;AAIrE,MAAM,WAAW,gBAAgB;IAC/B,kEAAkE;IAClE,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;IAEpB;;;;OAIG;IACH,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAE9B;;;;;OAKG;IACH,QAAQ,CAAC,KAAK,EAAE,WAAW,CAAC;IAE5B;;;;OAIG;IACH,QAAQ,CAAC,QAAQ,EAAE,QAAQ,CAAC;IAE5B;;;;OAIG;IACH,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,CAAC;IAE7B;;;;OAIG;IACH,QAAQ,CAAC,IAAI,CAAC,EAAE,MAAM,CAAC;IAEvB;;;;OAIG;IACH,QAAQ,CAAC,SAAS,CAAC,EAAE,MAAM,CAAC;IAE5B;;;;;OAKG;IACH,QAAQ,CAAC,MAAM,CAAC,EAAE,WAAW,CAAC;CAC/B;AAED;;;;;;;GAOG;AACH,wBAAgB,SAAS,CAAC,IAAI,EAAE,gBAAgB,GAAG,gBAAgB,CA8BlE"}

package/dist/types/lib/rag/index.d.ts

@@ -0,0 +1,8 @@
+/**
+ * RAG — retrieval-augmented generation as a context-engineering
+ * flavor. ONE factory + ONE seeding helper. Composes over the memory
+ * subsystem (semantic + top-K + strict threshold).
+ */
+export { defineRAG, type DefineRAGOptions } from './defineRAG.js';
+export { indexDocuments, type IndexDocumentsOptions, type RagDocument } from './indexDocuments.js';
+//# sourceMappingURL=index.d.ts.map

package/dist/types/lib/rag/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../../src/lib/rag/index.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AACH,OAAO,EAAE,SAAS,EAAE,KAAK,gBAAgB,EAAE,MAAM,gBAAgB,CAAC;AAClE,OAAO,EAAE,cAAc,EAAE,KAAK,qBAAqB,EAAE,KAAK,WAAW,EAAE,MAAM,qBAAqB,CAAC"}

package/dist/types/lib/rag/indexDocuments.d.ts

@@ -0,0 +1,77 @@
+/**
+ * indexDocuments — seed a vector-capable MemoryStore with documents.
+ *
+ * Embeds each document, builds a `MemoryEntry<{content, metadata?}>`,
+ * batches into `store.putMany()`. Used at application startup to
+ * populate a RAG store before the first agent run.
+ *
+ * Pattern: Bulk-write helper. Not a flowchart stage — it runs once
+ *          at boot, not per-iteration.
+ * Role:    Layer-3 RAG pipeline starter. Pairs with `defineRAG()`
+ *          which only does the read side.
+ * Emits:   N/A — startup-time batch write, not part of the agent run.
+ *
+ * @example
+ * ```ts
+ * import { InMemoryStore, mockEmbedder, indexDocuments, defineRAG } from 'agentfootprint';
+ *
+ * const store = new InMemoryStore();
+ * const embedder = mockEmbedder();
+ *
+ * await indexDocuments(store, embedder, [
+ *   { id: 'doc1', content: 'Refunds processed within 3 business days.' },
+ *   { id: 'doc2', content: 'Pro plan: $20/mo, includes priority support.', metadata: { tier: 'pro' } },
+ *   { id: 'doc3', content: 'Free plan: limited to 100 calls/month.' },
+ * ]);
+ *
+ * const docs = defineRAG({ id: 'product-docs', store, embedder });
+ * agent.rag(docs);
+ * ```
+ */
+import type { Embedder } from '../../memory/embedding/index.js';
+import type { MemoryStore } from '../../memory/store/index.js';
+import type { MemoryIdentity } from '../../memory/identity/index.js';
+/** A document to index. `id` must be unique within the store + identity. */
+export interface RagDocument {
+    readonly id: string;
+    readonly content: string;
+    readonly metadata?: Readonly<Record<string, unknown>>;
+}
+export interface IndexDocumentsOptions {
+    /**
+     * Identity scope to write under. Default: a single shared
+     * `{ conversationId: '_global' }` namespace, suitable for app-wide
+     * corpora. Override for per-tenant document partitions.
+     */
+    readonly identity?: MemoryIdentity;
+    /**
+     * Stable id of the embedder. Stored on each entry so a future
+     * embedder swap doesn't silently mix similarity scores. Default:
+     * `'default-embedder'` — pass an explicit id when you may rotate
+     * embedders.
+     */
+    readonly embedderId?: string;
+    /**
+     * Optional tier tag to attach to indexed entries (`'hot'` /
+     * `'warm'` / `'cold'`). Useful when read-side `defineRAG` should
+     * filter to a subset of the corpus.
+     */
+    readonly tier?: 'hot' | 'warm' | 'cold';
+    /**
+     * Optional TTL in milliseconds from indexing time. Useful for
+     * compliance retention windows (e.g., re-index quarterly).
+     */
+    readonly ttlMs?: number;
+    /**
+     * Optional abort signal — embedders making network calls thread
+     * this through to abort batch indexing on shutdown / timeout.
+     */
+    readonly signal?: AbortSignal;
+}
+/**
+ * Embed + persist documents. Returns the count actually indexed
+ * (skips duplicates if the store rejects them). Throws on embedder
+ * failure or store error — fail loud at startup is desirable.
+ */
+export declare function indexDocuments(store: MemoryStore, embedder: Embedder, documents: readonly RagDocument[], options?: IndexDocumentsOptions): Promise<number>;
+//# sourceMappingURL=indexDocuments.d.ts.map

package/dist/types/lib/rag/indexDocuments.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"indexDocuments.d.ts","sourceRoot":"","sources":["../../../../src/lib/rag/indexDocuments.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6BG;AAEH,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,iCAAiC,CAAC;AAEhE,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,6BAA6B,CAAC;AAC/D,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,gCAAgC,CAAC;AAErE,4EAA4E;AAC5E,MAAM,WAAW,WAAW;IAC1B,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;IACpB,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;IACzB,QAAQ,CAAC,QAAQ,CAAC,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC;CACvD;AAED,MAAM,WAAW,qBAAqB;IACpC;;;;OAIG;IACH,QAAQ,CAAC,QAAQ,CAAC,EAAE,cAAc,CAAC;IAEnC;;;;;OAKG;IACH,QAAQ,CAAC,UAAU,CAAC,EAAE,MAAM,CAAC;IAE7B;;;;OAIG;IACH,QAAQ,CAAC,IAAI,CAAC,EAAE,KAAK,GAAG,MAAM,GAAG,MAAM,CAAC;IAExC;;;OAGG;IACH,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IAExB;;;OAGG;IACH,QAAQ,CAAC,MAAM,CAAC,EAAE,WAAW,CAAC;CAC/B;AAID;;;;GAIG;AACH,wBAAsB,cAAc,CAClC,KAAK,EAAE,WAAW,EAClB,QAAQ,EAAE,QAAQ,EAClB,SAAS,EAAE,SAAS,WAAW,EAAE,EACjC,OAAO,GAAE,qBAA0B,GAClC,OAAO,CAAC,MAAM,CAAC,CAwCjB"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentfootprint",
-  "version": "2.0.1",
+  "version": "2.1.0",
   "description": "The explainable agent framework — build AI agents you can explain, audit, and trust. Built on footprintjs.",
   "license": "MIT",
   "author": "Sanjay Krishna Anbalagan",