@oribos/rag

Semantic recall over message history — the RAG capability package. It wraps a MemoryStore so that every message Memory.save persists is indexed with your embedding model, and the window Memory.recall returns is merged with older, semantically related messages. One seam, no core changes:

import { Memory } from '@oribos/core/memory';
import { createSemanticMemoryStore } from '@oribos/rag';
import { openai } from '@ai-sdk/openai';

const memory = new Memory({
  storage: createSemanticMemoryStore({
    embedder: openai.textEmbeddingModel('text-embedding-3-small'),
  }),
  lastMessages: 10,
});

How it works

• Save indexes. createSemanticMemoryStore wraps any MemoryStore (absent = the core's in-memory default). Each persisted message's text (system / user / assistant text parts) goes to the embedder and into a vector index — inline, within the save call. Messages persist first; an embedder failure rejects the save explicitly.
• Recall merges. A cursorless window read embeds the window's last user message and merges positively-similar older messages back into the result, in chronological position — the recall window forgets less. Paging with before passes through untouched; an empty or textless window spends no embedding. topK defaults to 3.
• Pins kept. Messages are immutable (the index is a derived cache of full stored messages), recall stays the single query entry, deleteThread cascades to the index, and the working-memory capability of the wrapped store passes through.

The embedder follows the framework's model contract pattern: a vendored structural subset of the AI SDK provider specification's embedding model interface (v4). Any embedding model instance from a provider package satisfies it — no adapter, no registry. The vector index is the package's own port: createInMemoryVectorIndex() ships as the zero-dependency default; an external vector database rides as your VectorIndex implementation, so its dependency travels with this package (or your adapter), never the core. Unconfigured, nothing exists — opt-in, zero cost, no resident pipeline.

See docs/architecture/memory.md (「语义召回」) and examples/semantic-recall.