agents-inc/api-vector-db-chroma

Chroma Patterns

Quick Guide: Use chromadb (v3.x) with @chroma-core/default-embed for automatic embedding. Chroma auto-embeds documents if no embeddings are provided -- just pass documents and ids to collection.add(). Use where for metadata filtering and whereDocument for document content filtering ($contains, $regex). Default distance metric is l2 (Euclidean); use cosine for most embedding models via configuration: { hnsw: { space: "cosine" } }. Query results return nested arrays (ids: string[][]) because queries are batched -- always access results.ids[0] for a single query. Include only the fields you need via the include parameter to reduce payload size.

---

<critical_requirements>

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 install @chroma-core/default-embed alongside chromadb -- the default embedding function ships as a separate package since v3)

(You MUST access query results as nested arrays -- results.ids[0], results.documents[0] -- because Chroma batches queries and returns string[][] not string[])

(You MUST use the configuration parameter for HNSW settings -- the legacy metadata: { "hnsw:space": "cosine" } approach is deprecated)

(You MUST use flat metadata values only (string, number, boolean, typed arrays) -- nested objects are not supported and will be rejected)

</critical_requirements>

---

Examples

• Core Patterns -- Client setup, collection management, add, query, get, update, upsert, delete
• Metadata Filtering -- Filter operators, compound filters, document content filters, whereDocument
• Embedding Functions -- Default, OpenAI, custom embedding functions, provider packages

Additional resources:

• reference.md -- API quick reference, filter operators, include options, limits, production checklist

---

Auto-detection: Chroma, chromadb, ChromaClient, CloudClient, createCollection, getOrCreateCollection, collection.add, collection.query, collection.get, collection.upsert, queryTexts, queryEmbeddings, nResults, whereDocument, $contains, @chroma-core/default-embed, @chroma-core/openai, EmbeddingFunction, vector database, semantic search, embedding, RAG retrieval, hnsw:space

When to use:

• Semantic search over document embeddings (RAG retrieval)
• Rapid prototyping with automatic embedding generation (no external embedding pipeline needed)
• Metadata-filtered vector search with compound logical operators
• Document content filtering with $contains and $regex
• Local development with in-process or Docker-based Chroma server

Key patterns covered:

• Client setup (HTTP, Cloud, Docker)
• Collection management (create, get, delete, configure HNSW)
• Document CRUD with automatic embedding (add, query, get, update, upsert, delete)
• Metadata filtering (where) with comparison, set, array, and logical operators
• Document content filtering (whereDocument) with $contains and $regex
• Embedding function configuration (default, OpenAI, custom)
• Query result handling (nested array structure, include options)

When NOT to use:

• Full-text search with complex boolean ranking (use a dedicated search engine)
• Relational data with joins and transactions (use a relational database)
• Multi-modal image+text embeddings in TypeScript (currently Python-only in Chroma)
• High-scale production with millions of vectors and strict SLAs (evaluate managed vector databases)

---

<philosophy>

Philosophy

Chroma is a lightweight, developer-friendly embedding database designed for rapid prototyping and production RAG applications. The core principle: pass documents in, get relevant results out -- Chroma handles embedding automatically.

Core principles:

1. Documents first, vectors optional -- Unlike most vector databases, Chroma can embed documents automatically using a configured embedding function. You never need to manage embeddings directly unless you want to.
2. Collections are self-contained -- Each collection has its own embedding function, distance metric, and HNSW configuration. No global index management needed.
3. Metadata is for filtering, documents are for content -- Use where for structured metadata filters and whereDocument for full-text content filters. Both can be combined in a single query.
4. Batteries included -- The default embedding function (all-MiniLM-L6-v2 via @chroma-core/default-embed) works out of the box for English text. Swap to OpenAI, Cohere, or any provider with a single package change.
5. Query results are batched -- Chroma supports multiple queries in a single call. Results are always nested arrays (string[][]), even for single queries. Always access [0] for the first query's results.

</philosophy>

---

<patterns>

Core Patterns

Pattern 1: Client Initialization

Always pass the server URL explicitly from an environment variable -- never rely on the implicit http://localhost:8000 default. See examples/core.md for HTTP, Cloud, and token-authenticated client examples.

const chromaUrl = process.env.CHROMA_URL;
if (!chromaUrl) throw new Error("CHROMA_URL environment variable is required");
return new ChromaClient({ path: chromaUrl });

---

Pattern 2: Collection with Distance Metric

Use the configuration parameter for HNSW settings -- never the deprecated metadata: { "hnsw:space": "cosine" } approach. See examples/core.md.

const collection = await client.createCollection({
  name: COLLECTION_NAME,
  configuration: { hnsw: { space: "cosine" } },
});

---

Pattern 3: Add Documents with Automatic Embedding

Pass documents and ids -- Chroma embeds automatically. Either documents or embeddings must be provided; metadata alone is insufficient. See examples/core.md.

await collection.add({
  ids: articles.map((a) => a.id),
  documents: articles.map((a) => a.text),
  metadatas: articles.map((a) => ({ category: a.category })),
});

---

Pattern 4: Query with Metadata and Document Filters

Combine where (metadata) and whereDocument (content) filters. Results are nested arrays -- always access [0] for single-query results. See examples/metadata-filtering.md for all operators.

const results = await collection.query({
  queryTexts: ["machine learning fundamentals"],
  nResults: N_RESULTS,
  where: {
    $and: [{ category: { $eq: "tutorial" } }, { year: { $gte: 2023 } }],
  },
  whereDocument: { $contains: "neural network" },
  include: ["documents", "metadatas", "distances"],
});
// results.ids[0] -- nested array, access [0] for first query

---

Pattern 5: Get with Pagination

Use get() with limit/offset for non-similarity retrieval. Unlike query(), get() returns flat arrays. See examples/core.md.

const results = await collection.get({
  where: { category: { $eq: category } },
  limit: PAGE_SIZE,
  offset,
  include: ["documents", "metadatas"],
});

---

Pattern 6: Upsert for Idempotent Operations

Use upsert() instead of add() for create-or-update semantics -- safer for pipelines that run multiple times. See examples/core.md.

await collection.upsert({
  ids: docs.map((d) => d.id),
  documents: docs.map((d) => d.text),
  metadatas: docs.map((d) => d.metadata),
});

</patterns>

---

<decision_framework>

Decision Framework

Which Distance Metric?

Which distance metric should I use?
|-- Using embeddings from a language model? -> cosine (normalized, most common)
|-- Need dot product similarity? -> ip (inner product)
|-- Comparing raw feature vectors? -> l2 (Euclidean, Chroma default)
'-- Unsure? -> cosine (safe default for most embedding models)

Which Embedding Function?

Which embedding function should I use?
|-- Quick prototyping, English text? -> @chroma-core/default-embed (all-MiniLM-L6-v2, runs locally)
|-- Need high-quality embeddings? -> @chroma-core/openai (text-embedding-3-small)
|-- Have your own embedding pipeline? -> Pass embeddings directly (skip embedding function)
|-- Need custom model? -> Implement EmbeddingFunction interface
'-- Want all providers? -> npm install @chroma-core/all

Where vs WhereDocument?

How should I filter results?
|-- Structured attributes (category, year, status)? -> where (metadata filter)
|-- Full-text content search? -> whereDocument ($contains, $regex)
|-- Both? -> Combine where + whereDocument in same query
'-- Need exact match on specific IDs? -> get({ ids: [...] })

ChromaClient vs CloudClient?

Which client should I use?
|-- Local development or self-hosted? -> ChromaClient({ path: "http://localhost:8000" })
|-- Chroma Cloud (managed)? -> CloudClient({ apiKey, tenant, database })
|-- Docker deployment? -> ChromaClient with Docker host URL
'-- Testing? -> ChromaClient against local Docker container

</decision_framework>

---

<red_flags>

RED FLAGS

High Priority Issues:

• Accessing query results as flat arrays instead of nested -- results.ids is string[][], not string[]; always use results.ids[0] for single-query results
• Missing @chroma-core/default-embed package -- since v3, the default embedding function ships separately; npm install chromadb @chroma-core/default-embed
• Using deprecated metadata: { "hnsw:space": "cosine" } for HNSW config -- use configuration: { hnsw: { space: "cosine" } } instead
• Nested objects in metadata -- Chroma only supports flat key-value metadata; nested objects are rejected

Medium Priority Issues:

• Not specifying include in queries -- default includes vary (query returns documents, metadatas, distances; get returns documents, metadatas); explicitly set include for clarity and to control payload size
• Using l2 (default) when cosine is appropriate -- most embedding models are normalized for cosine similarity; l2 may produce worse results
• Calling add() without documents or embeddings -- at least one must be provided; metadata alone is insufficient
• Not handling empty results -- results.ids[0] may be an empty array; check length before processing

Common Mistakes:

• Passing queryEmbeddings AND queryTexts together -- use one or the other, not both
• Expecting update() to create missing records -- update() silently ignores non-existent IDs; use upsert() for create-or-update semantics
• Calling delete() with no arguments -- deletes nothing (not everything); pass ids or where to target specific records
• Using $gt/$lt on string metadata -- comparison operators only work on numeric values (int or float)

Gotchas & Edge Cases:

• HNSW configuration (space, ef_construction, max_neighbors) cannot be changed after collection creation -- you must delete and recreate the collection
• collection.count() returns total records in the collection, not filtered counts -- there is no filtered count API
• peek() returns the first limit items (default 10) in insertion order, not by relevance -- useful for debugging, not querying
• $contains in whereDocument is case-sensitive -- searching for "Neural" will not match "neural"
• $regex in whereDocument uses full regex syntax but can be slow on large collections
• Array metadata values (string[], number[]) must be homogeneous -- mixing types within an array is rejected
• Metadata keys are case-sensitive -- Category and category are different fields
• The nResults default is 10 if not specified in query()
• Multimodal embedding (images + text) is currently Python-only -- TypeScript support is not yet available

</red_flags>

---

<critical_reminders>

CRITICAL REMINDERS

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

(You MUST install @chroma-core/default-embed alongside chromadb -- the default embedding function ships as a separate package since v3)

(You MUST access query results as nested arrays -- results.ids[0], results.documents[0] -- because Chroma batches queries and returns string[][] not string[])

(You MUST use the configuration parameter for HNSW settings -- the legacy metadata: { "hnsw:space": "cosine" } approach is deprecated)

(You MUST use flat metadata values only (string, number, boolean, typed arrays) -- nested objects are not supported and will be rejected)

Failure to follow these rules will cause embedding failures, incorrect result access, deprecated configuration warnings, and rejected metadata.

</critical_reminders>