Vector stores and chat history

Every IntelliNode vector store has the same five methods, so you can start with an in-memory store and move to a database later without changing the rest of your code. The Assistant uses these stores for documents and memory, and you can also use them on their own.

The shared interface

const { QdrantVectorStore } = require('intellinode');

const store = new QdrantVectorStore({
  url: 'http://localhost:6333',
  collection: 'docs',
  embedder: { provider: 'openai', apiKey: process.env.OPENAI_API_KEY },
});

await store.addDocuments([{ id: 'a', text: 'IntelliNode supports Gemini.', metadata: { lang: 'en' } }]);
const hits = await store.search('Which models are supported?', 3, { lang: 'en' });
// [{ id, score, text, metadata }]

Method	What it does
addDocuments(docs)	Embeds the text with the store's embedder and saves it under your ids.
upsert(items)	Saves vectors you made yourself: { id, vector, text, metadata }.
search(text, topK, filter)	Finds the closest documents to a question.
query({ text or vector, topK, filter })	The same search, with an object.
delete(ids)	Removes documents by id.

score is a similarity, so higher means closer. filter matches metadata values exactly, and an array means "one of". For anything more, pass the database's own filter as nativeFilter.

Pick an embedder

The embedder turns text into vectors:

{ provider: 'openai', apiKey }                                   // text-embedding-3-small
{ provider: 'gemini', apiKey }                                   // 3072 numbers
{ provider: 'vertex', apiKey, dimensions: 768 }                  // gemini-embedding-001 on Vertex AI
{ provider: 'cohere', apiKey }
{ provider: 'ollama', model: 'nomic-embed-text', options: { baseUrl: 'http://localhost:11434/v1' } }
async (texts) => myVectors(texts)                                 // your own function

Keep the same embedder for the life of a store. A different model or dimensions makes the saved vectors useless.

Pick a store

Class	Use it for	Needs
MemoryVectorStore({ path })	Local apps, tests and the browser, up to a few thousand chunks	Nothing. path saves it to a JSON file.
PineconeVectorStore({ apiKey, indexHost })	Managed vector search	An index with the right dimension
QdrantVectorStore({ url, apiKey, collection })	Self-hosted or Qdrant Cloud	A Qdrant server
ChromaVectorStore({ url, collection })	Local development	chroma run or Chroma Cloud
WeaviateVectorStore({ url, apiKey, className })	Weaviate 1.2x and later	A capitalized class name
MilvusVectorStore({ url, token, collection, dimension })	Milvus or Zilliz	Milvus 2.5 or later
ElasticsearchVectorStore({ url, apiKey, index, dimension })	An existing Elasticsearch cluster	Elasticsearch 8 or later
PgVectorStore({ client, dimension })	Postgres, AlloyDB, Cloud SQL, Supabase, Neon	npm i pg and the pgvector extension
MongoDBAtlasVectorStore({ collection })	MongoDB Atlas	npm i mongodb and an Atlas vector index
FirestoreVectorStore({ projectId, collection })	Data next to your Firestore app data	Google Cloud OAuth and a vector index
VertexRAGStore({ projectId, location, corpus })	Google parses, chunks and embeds your files	Google Cloud OAuth
VertexVectorSearchStore({ projectId, collection })	Vertex AI Vector Search 2.0	Google Cloud OAuth

PgVectorStore creates its table and index on first use, and Qdrant creates its collection on the first write.

Google Cloud stores

Firestore, RAG Engine and Vector Search don't accept API keys. For local development, sign in once:

gcloud auth application-default login

On Cloud Run, GKE or Compute Engine the token comes from the environment. Anywhere else, point GOOGLE_APPLICATION_CREDENTIALS at a service account key file, or pass accessToken.

Firestore also needs a vector index. store.indexCommand(768) prints the gcloud command that creates it.

Chat history

The Assistant saves conversations through one of these:

Class	Where it saves	Good for
MemoryChatHistory()	Process memory, lost on restart	Tests. It is the default.
FileChatHistory({ dir })	One JSON file per conversation	Local apps and desktop tools
FirestoreChatHistory({ projectId, collection })	Firestore	Servers with many users

With Firestore for both history and memory, a user's conversations stay in your Google Cloud project:

const { Assistant, FirestoreChatHistory, FirestoreVectorStore } = require('intellinode');

const embedder = { provider: 'vertex', apiKey: process.env.VERTEX_API_KEY, dimensions: 768 };
const assistant = new Assistant({
  provider: 'vertex',
  apiKey: process.env.VERTEX_API_KEY,
  history: new FirestoreChatHistory({ projectId }),
  memory: new FirestoreVectorStore({ projectId, collection: 'memories', embedder }),
});

For another database, such as Redis or DynamoDB, extend ChatHistory and write its seven methods: getMessages, addMessages, getConversation, saveConversation, listConversations, deleteConversation and deleteLastMessages.