API Vector DB Weaviate

Weaviate Patterns

Quick Guide: Use Weaviate for semantic search and RAG applications. Use weaviate-client (v3.x) as the TypeScript client -- it uses gRPC for performance and provides full type safety with generics. Connect via connectToWeaviateCloud() for managed instances or connectToLocal() for Docker. Collections are the central abstraction -- configure vectorizers at collection level, not per-query. Use collection.query.* for search, collection.generate.* for RAG, and collection.data.* for CRUD. Always call client.close() when done. Increase query timeout to 60s+ when using generative search. The v3 client does NOT support browsers or Embedded Weaviate.

CRITICAL: Before Using This Skill

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST call client.close() when done with the Weaviate client -- it maintains persistent gRPC connections that will leak if not closed)

(You MUST configure vectorizers at the COLLECTION level during client.collections.create() -- you cannot add a vectorizer after creation, only add new named vectors)

(You MUST use a SEPARATE client.collections.use() call with .withTenant() for multi-tenant queries -- queries without tenant context on multi-tenant collections will fail)

(You MUST increase query timeout to 60+ seconds when using generate.* (RAG) submodule -- generative model calls are slow and the default timeout causes failures)

Examples

• Core Patterns -- Connection, collection setup, object CRUD, basic search
• Search & Filtering -- nearText, nearVector, hybrid, bm25, filters, generative search (RAG)
• Multi-Tenancy & Batch -- Tenant management, batch imports, cross-references

Additional resources:

• reference.md -- API cheat sheet, vectorizer comparison, data types, decision frameworks

Auto-detection: Weaviate, weaviate-client, connectToWeaviateCloud, connectToLocal, nearText, nearVector, hybrid search, bm25, vector database, semantic search, RAG, generative search, generate.nearText, insertMany, vectorizer, text2vec, multi-tenancy, withTenant, collection.query, collection.generate, collection.data

When to use:

• Semantic search over text, images, or multimodal data
• Retrieval Augmented Generation (RAG) with built-in generative search
• Hybrid search combining vector similarity and keyword (BM25) ranking
• Multi-tenant applications needing isolated vector stores per customer
• Applications requiring built-in vectorization (no external embedding pipeline)
• Real-time similarity search with filtering on structured properties

Key patterns covered:

• weaviate-client v3 connection setup and configuration
• Collection management with vectorizer modules (text2vec-openai, text2vec-cohere, etc.)
• Object CRUD (insert, insertMany, update, replace, deleteById, deleteMany)
• Search types (nearText, nearVector, hybrid, bm25, fetchObjects)
• Filtering with operators (equal, greaterThan, like, containsAny, and/or/not)
• Generative search (RAG) with singlePrompt and groupedTask
• Multi-tenancy with tenant lifecycle management
• Batch imports with insertMany and error handling
• Cross-references between collections
• Named vectors for multi-vector collections

When NOT to use:

• Relational data with complex joins (use a relational database)
• Full-text search without vector component (use a dedicated search engine)
• Key-value caching (use a key-value store)
• Time-series data (use a time-series database)
• Graph traversal queries (use a graph database)
• Browser-side applications (v3 client is Node.js only)

Philosophy

Weaviate is a vector database that stores data objects alongside their vector embeddings. The core principle: configure once at the collection level, then query with simple method calls.

Core principles:

1. Collection-centric design -- All configuration (vectorizer, generative model, reranker, properties) is set at collection creation. Queries operate on collection objects obtained via client.collections.use().
2. Built-in vectorization -- Weaviate can vectorize data automatically using configured modules (text2vec-openai, text2vec-cohere, etc.). You don't need an external embedding pipeline unless you want one.
3. Search is a spectrum -- Use nearText for semantic similarity, bm25 for keyword matching, hybrid for a weighted combination. The alpha parameter controls the vector-vs-keyword balance in hybrid search.
4. RAG is a search mode, not a separate system -- Switch from collection.query.nearText() to collection.generate.nearText() to add LLM generation on top of search results.
5. Filters are additive -- Filters narrow results after vector/keyword retrieval. Combine with Filters.and() and Filters.or() for complex conditions.

Core Patterns

Pattern 1: Connection Setup

Connect to Weaviate Cloud or local Docker instance. Always close the client when done. See examples/core.md for full examples.

// Good Example -- Cloud connection with API key headers
import weaviate from "weaviate-client";

const QUERY_TIMEOUT_SECONDS = 30;
const INSERT_TIMEOUT_SECONDS = 120;

async function createWeaviateClient() {
  const client = await weaviate.connectToWeaviateCloud(
    process.env.WEAVIATE_URL!,
    {
      authCredentials: new weaviate.ApiKey(process.env.WEAVIATE_API_KEY!),
      headers: {
        "X-OpenAI-Api-Key": process.env.OPENAI_API_KEY!,
      },
      timeout: {
        query: QUERY_TIMEOUT_SECONDS,
        insert: INSERT_TIMEOUT_SECONDS,
      },
    },
  );
  return client;
}

export { createWeaviateClient };

Why good: Environment variables for credentials, explicit timeouts, API key headers for vectorizer modules

// Bad Example -- Missing cleanup, no timeout config
import weaviate from "weaviate-client";
const client = await weaviate.connectToLocal();
// No client.close() -- gRPC connections leak
// No timeout config -- generative queries will timeout

Why bad: Missing client.close() leaks gRPC connections, default timeout too short for RAG queries

Pattern 2: Collection with Vectorizer

Configure vectorizer and properties at creation time. See examples/core.md for named vectors and advanced configuration.

// Good Example -- Collection with vectorizer and generative model
import { vectors, dataType, generative } from "weaviate-client";

await client.collections.create({
  name: "Article",
  vectorizers: vectors.text2VecOpenAI({
    model: "text-embedding-3-small",
  }),
  generative: generative.openAI({
    model: "gpt-4o",
  }),
  properties: [
    { name: "title", dataType: dataType.TEXT },
    { name: "body", dataType: dataType.TEXT },
    { name: "category", dataType: dataType.TEXT },
    { name: "publishedAt", dataType: dataType.DATE },
  ],
});

Why good: Vectorizer and generative model configured at collection level, typed properties with explicit data types

// Bad Example -- Trying to add vectorizer after creation
await client.collections.create({ name: "Article" });
// No way to add a vectorizer to an existing collection without named vectors
// Must delete and recreate, or use addVector() for named vectors only

Why bad: Vectorizer must be set at creation time; cannot be added to an existing default vector after the fact

Pattern 3: Hybrid Search with Filters

Combine vector and keyword search with property filters. See examples/search.md for all search types.

// Good Example -- Hybrid search with filter
import { Filters } from "weaviate-client";

const articles = client.collections.use("Article");
const SEARCH_LIMIT = 10;
const HYBRID_ALPHA = 0.75; // Favor vector search

const result = await articles.query.hybrid("machine learning trends", {
  alpha: HYBRID_ALPHA,
  limit: SEARCH_LIMIT,
  filters: Filters.and(
    articles.filter.byProperty("category").equal("technology"),
    articles.filter
      .byProperty("publishedAt")
      .greaterThan(new Date("2024-01-01")),
  ),
  returnMetadata: ["score", "explainScore"],
});

for (const obj of result.objects) {
  console.log(obj.properties.title, obj.metadata?.score);
}

Why good: Named constants for limits and alpha, combined filter with Filters.and(), metadata for debugging relevance

Pattern 4: Generative Search (RAG)

Switch from query.* to generate.* for RAG. See examples/search.md for singlePrompt and groupedTask patterns.

// Good Example -- RAG with single prompt per result
const articles = client.collections.use("Article");
const RAG_RESULT_LIMIT = 5;

const result = await articles.generate.nearText(
  "climate change policy",
  {
    singlePrompt: "Summarize this article in one sentence: {title} - {body}",
  },
  {
    limit: RAG_RESULT_LIMIT,
    returnMetadata: ["distance"],
  },
);

for (const obj of result.objects) {
  console.log("Source:", obj.properties.title);
  console.log("Generated:", obj.generative?.text);
}

Why good: Uses property interpolation {title} in prompt, accesses generated text via obj.generative?.text

// Bad Example -- Using query instead of generate for RAG
const result = await articles.query.nearText("climate change", { limit: 5 });
// Then manually calling OpenAI API with results
// Weaviate does this natively with generate.*

Why bad: Misses Weaviate's built-in RAG -- extra network hops, no automatic prompt interpolation

Decision Framework

Which Search Type?

What kind of search do I need?
├─ Natural language query, semantic meaning? -> nearText (requires vectorizer module)
├─ Have pre-computed vector embedding? -> nearVector
├─ Exact keyword matching? -> bm25
├─ Both semantic and keyword relevance? -> hybrid (alpha controls blend)
├─ Just list/filter objects without search? -> fetchObjects
└─ Search + LLM generation? -> generate.nearText / generate.hybrid

Which Vectorizer?

Which vectorizer module should I use?
├─ OpenAI models (text-embedding-3-small/large)? -> text2VecOpenAI
├─ Cohere models (embed-v3)? -> text2VecCohere
├─ Self-hosted models? -> text2VecOllama or text2VecTransformers
├─ Bring your own embeddings? -> none (use selfProvided for named vectors)
├─ Multimodal (images + text)? -> multi2VecClip or multi2VecBind
└─ Multiple embedding strategies? -> Named vectors (array of vectorizers)

Single vs Named Vectors?

How many vector representations do I need?
├─ One embedding per object (most common)? -> Single default vectorizer
├─ Different embeddings for different properties? -> Named vectors
├─ Mix of auto-vectorized and self-provided? -> Named vectors with selfProvided
└─ Different models for different search use cases? -> Named vectors

When to Use Multi-Tenancy?

Do I need data isolation?
├─ Each customer/user needs isolated data? -> Enable multi-tenancy
├─ Shared dataset, filter by user? -> Single tenant with filters
├─ Need to offload inactive tenants? -> Multi-tenancy with tenant states
└─ Small number of distinct datasets? -> Separate collections may be simpler

RED FLAGS

High Priority Issues:

• Missing client.close() -- gRPC connections persist and leak memory/file descriptors
• Trying to add a default vectorizer after collection creation -- vectorizer must be configured in create(). Only named vectors can be added later with config.addVector()
• Querying a multi-tenant collection without .withTenant() -- all operations fail with an error
• Using default query timeout with generate.* -- generative calls need 60+ seconds; default is often too short

Medium Priority Issues:

• Using replace() when update() is intended -- replace deletes all properties not included in the call; update merges
• Not checking insertMany response for errors -- partial failures are silent; check response.hasErrors and response.errors
• Passing alpha: 1.0 to hybrid search -- equivalent to pure vector search; use nearText instead for clarity
• Not specifying targetVector with named vectors -- queries default to the first vector, which may not be the intended one

Common Mistakes:

• Using v2 class-based API (client.schema.classCreator()) with v3 client -- the API is completely different; v3 uses client.collections.create()
• Forgetting to pass API key headers for vectorizer modules -- X-OpenAI-Api-Key, X-Cohere-Api-Key etc. must be in connection headers
• Using connectToWCS() (deprecated) instead of connectToWeaviateCloud()
• Adding a property after data import without reindexing -- pre-existing objects won't have that property indexed

Gotchas & Edge Cases:

• insertMany uses server-side batching but the TS client does NOT have a streaming batch API -- for very large imports (100K+), chunk into batches of 100-1000 objects
• Filters.and() and Filters.or() take a flat list of filter conditions, NOT nested arrays -- Filters.and(a, b, c) not Filters.and([a, b, c])
• fetchObjects() without limit returns 25 objects by default (server-side default), not all objects
• Property names in Weaviate must start with a lowercase letter -- the client silently lowercases the first character
• distance metadata varies by vector distance metric -- cosine distance range [0, 2], not [0, 1]
• deleteMany has a server-side maximum of 10,000 objects per call (configurable via QUERY_MAXIMUM_RESULTS)
• Weaviate auto-detects property types on first insert if not defined in the schema -- this can cause type mismatches if first object has atypical data
• fetchObjectById returns null for non-existent IDs, not an empty object -- always check for null before accessing properties
• Cross-references in multi-tenant collections can only reference objects in the same tenant or in non-multi-tenant collections

CRITICAL REMINDERS

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST call client.close() when done with the Weaviate client -- it maintains persistent gRPC connections that will leak if not closed)

(You MUST configure vectorizers at the COLLECTION level during client.collections.create() -- you cannot add a vectorizer after creation, only add new named vectors)

(You MUST use a SEPARATE client.collections.use() call with .withTenant() for multi-tenant queries -- queries without tenant context on multi-tenant collections will fail)

(You MUST increase query timeout to 60+ seconds when using generate.* (RAG) submodule -- generative model calls are slow and the default timeout causes failures)

Failure to follow these rules will cause connection leaks, missing vectorization, multi-tenant query failures, and RAG timeouts.