lossless-monorepo/agent-skills/chroma-agent-skills/src/chroma-local/skill

Instructions

Determine these before writing code. Prefer discovering them from the repo and the user request. Ask only when the choice materially changes the implementation.

1. Runtime shape

   • Are they connecting to a running local server, embedding Chroma into tests, or setting up local development from scratch?
   • Decide whether they need chroma run, a Docker or service command, HttpClient or ChromaClient, or Python EphemeralClient.

2. Persistence

   • Persistent local data: choose an intentional data path.
   • Disposable test data: use defaults or a temp directory.

3. Embedding model

   • Reuse the app’s existing embedding provider when possible.
   • Otherwise default to @chroma-core/default-embed in TypeScript or the standard local default in Python.
   • If the user explicitly wants OpenAI embeddings in TypeScript, install and use @chroma-core/openai.

4. Indexed data shape

   • Determine what is being indexed, how it should be chunked, and what metadata is needed for filtering and updates.

Routing

• Existing local server

  • Confirm host and port before changing client code.
  • Validate the server is reachable before assuming collections are missing.

• Fresh local development

  • Add a local startup path such as chroma run or the repo’s existing Docker or service command.
  • Default to localhost:8000 unless the repo already uses another address.

• Python tests or disposable local workflows

  • Prefer EphemeralClient when persistence is unnecessary.
  • Call out that data is lost when the process exits.

• Persistent local development

  • Use a stable data path and make persistence explicit in code or config.
  • Do not silently switch between ephemeral and persistent modes.

• Search integration work

  • Use getOrCreateCollection() in TypeScript or get_or_create_collection() in Python.
  • Design document IDs and metadata so upserts and deletes are straightforward.
  • Batch writes when syncing large datasets.

Ask vs proceed

Ask first:

• Embedding model choice (cost and quality implications)
• Whether they need persistent local data
• How they are starting the local server
• Multi-tenant data isolation strategy

Proceed with sensible defaults:

• Use getOrCreateCollection() (TypeScript) / get_or_create_collection() (Python)
• Use cosine similarity (most common)
• Chunk size under 8KB
• Store source IDs in metadata for updates/deletes
• Use a local server on localhost:8000 unless the repo already configures another address or is using Python EphemeralClient

What to validate

• Correct client import (ChromaClient, HttpClient, or Client)
• Embedding function package is installed (TypeScript)
• Local server is reachable before assuming collections are missing
• Local path and persistence mode are intentional

Implementation notes

• Local Chroma is the right default for development, tests, and self-hosted deployments.
• OSS Chroma does not include Chroma Cloud-only features such as Schema() and Search().
• If the user asks for hybrid dense and sparse retrieval, treat that as a likely Chroma Cloud requirement unless the repo already implements an OSS workaround.
• For open source Chroma, dense retrieval with a single embedding function is the normal baseline.

Minimal patterns

Start a local Chroma server when the repo needs one:

chroma run

Default address: localhost:8000.

TypeScript local client:

import { ChromaClient } from 'chromadb';
import { DefaultEmbeddingFunction } from '@chroma-core/default-embed';

const client = new ChromaClient();

const embeddingFunction = new DefaultEmbeddingFunction();
const collection = await client.getOrCreateCollection({
  name: 'my_collection',
  embeddingFunction,
});

// Add documents
await collection.add({
  ids: ['doc1', 'doc2'],
  documents: ['First document text', 'Second document text'],
});

// Query
const results = await collection.query({
  queryTexts: ['search query'],
  nResults: 5,
});

Python local client:

import chromadb

client = chromadb.HttpClient(host="localhost", port=8000)

collection = client.get_or_create_collection(name="my_collection")

# Add documents
collection.add(
    ids=["doc1", "doc2"]   ,
    documents=["First document text", "Second document text"],
)

# Query
results = collection.query(
    query_texts=["search query"],
    n_results=5,
)

Learn More

Fetch Chroma’s llms.txt only when you need API or product details that are not already in the repo or this skill: https://docs.trychroma.com/llms.txt