Vector databases have revolutionized how we handle semantic search, but the real breakthrough comes when these systems can learn from their own mistakes. Chroma Context-1 represents a fundamental shift in search technology—training agents that continuously improve their search strategies through self-editing capabilities. This approach doesn't just find answers; it learns which search patterns work best and adapts accordingly.

What is Chroma Context-1?

Chroma Context-1 is Chroma's next-generation model designed specifically for training self-editing search agents. Unlike traditional retrieval systems that follow static rules, Context-1 enables agents to analyze their search results, identify gaps in relevance, and modify their approach in real-time.

The architecture combines vector embeddings with a feedback loop mechanism. When an agent performs a search, it evaluates the quality of results against expected outcomes. If results fall below a confidence threshold, the agent automatically refines its query or adjusts the search parameters without human intervention.

This self-editing capability means your search system gets smarter over time. Each interaction teaches the agent patterns that lead to better retrieval outcomes, reducing noise and improving precision across diverse query types.

Training Your First Self-Editing Agent

Training a self-editing search agent with Chroma Context-1 involves three core phases: initialization, feedback collection, and iterative refinement.

**Phase 1: Initialization**

Start by defining your corpus and establishing baseline performance metrics. Your agent needs a foundation of quality data to learn from effectively.

**Phase 2: Feedback Collection**

During this phase, the agent performs searches while you provide explicit or implicit signals about result quality. Explicit feedback includes user ratings or corrections, while implicit feedback comes from engagement metrics like click-through rates or session duration.

**Phase 3: Iterative Refinement**

The agent analyzes feedback patterns and adjusts its search strategy. Context-1's model weights update based on which query modifications led to improved results.

import chroma
from chroma.context_1 import SelfEditingAgent
Initialize the agent with your collection
collection = chroma.get_collection("knowledge_base")
agent = SelfEditingAgent(collection)
Train with sample interactions
training_data = [
    {"query": "machine learning basics", "feedback": 4},
    {"query": "neural network architecture", "feedback": 5},
    {"query": "deep learning concepts", "feedback": 3},
]
for interaction in training_data:
    agent.train(interaction["query"], interaction["feedback"])
Run inference
results = agent.search("ml fundamentals", auto_edit=True)

This simple example demonstrates how quickly you can deploy a self-improving search system. The auto_edit=True parameter enables automatic query refinement when initial results don't meet confidence thresholds.

Key Features and Implementation Tips

Chroma Context-1 offers several capabilities that make self-editing agents