Agent-ModernColBERT
by lightonai
150M late-interaction retriever optimized for agentic reasoning traces
lightonai/Agent-ModernColBERTmixpeek://text_extractor@v1/lighton_agent_moderncolbert_v1Overview
Agent-ModernColBERT is a 150M parameter late-interaction retrieval model from LightOn, specifically trained on agentic retrieval data where queries contain reasoning traces alongside the search intent. Built on ModernBERT architecture via PyLate, it achieves 72.53% accuracy on BrowseComp-Plus — exceeding configurations using GPT-5 + Qwen3-8B despite being 26x smaller than AgentIR-4B. This makes it uniquely suited for AI agent tool-use pipelines where the query is a chain-of-thought reasoning trace, not a clean user query.
Architecture
ModernBERT-based late-interaction model trained with PyLate on AgentIR data. Uses per-token 128-dim embeddings with MaxSim scoring, like ColBERT. The key innovation is training on reasoning trace + query pairs, so the model learns to extract search intent from noisy agentic context — function calls, intermediate thoughts, and partial conclusions.
Mixpeek SDK Integration
from mixpeek import Mixpeekmx = Mixpeek(api_key="YOUR_KEY")mx.ingest.documents(source="s3://knowledge-base/",collection="agent_kb",feature_extractors=[{"name": "text_embeddings","model": "lightonai/Agent-ModernColBERT","params": {"interaction": "late", "dim": 128}}])
Capabilities
- Retrieval from AI agent reasoning traces (not just clean queries)
- Late-interaction scoring for fine-grained token matching
- Tiny model footprint (150M) with outsized agentic performance
- Compatible with standard ColBERT indexing and serving
- Strong zero-shot transfer to general retrieval tasks
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| BrowseComp-Plus | Accuracy | 72.53% | Exceeds GPT-5 + Qwen3-8B setup |
| AgentIR | Retrieval Acc | Competitive with 4B models | At 150M params (26x smaller) |
Performance
Specification
Build a pipeline with Agent-ModernColBERT
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Open Studio