{
  "id": "colbertv2-0",
  "identity": {
    "family": "colbert",
    "display_name": "ColBERTv2",
    "hub_id": "colbert-ir/colbertv2.0",
    "api_name": null,
    "revision": null,
    "variants": [],
    "license": "MIT",
    "commercial_use": true,
    "commercial_use_note": "MIT on the Hugging Face repo; no use restrictions.",
    "release_date": "2021-12",
    "paper_url": "https://arxiv.org/abs/2112.01488",
    "repo_url": "https://github.com/stanford-futuredata/ColBERT"
  },
  "task": {
    "category": "embedding",
    "modalities_in": [
      "text"
    ],
    "modalities_out": [
      "multi_vector"
    ],
    "best_at": "Late-interaction text retrieval and reranking (MaxSim over one 128-d vector per token); strong on out-of-domain and term-heavy queries.",
    "worst_at": "Storage and cold reads: about 36 KB a 275-token document; English-centric BERT-base vocabulary; text only."
  },
  "io": {
    "input": {
      "text": {
        "max_tokens": 512,
        "query_marker": "[unused0] as [Q]",
        "document_marker": "[unused1] as [D]",
        "query_padding": "[MASK] to 32 positions"
      }
    },
    "output": {
      "vector": {
        "dims": 128,
        "distance": "cosine",
        "type": "multi_dense"
      }
    }
  },
  "mixpeek_wiring": {
    "extractors": [
      {
        "extractor": "colbert_extractor@v1",
        "selected_by": "fixed"
      }
    ],
    "feature_uris": [
      "mixpeek://colbert_extractor@v1/colbertv2_v1"
    ],
    "vector_indexes": [
      "colbert_extractor_v1_tokens"
    ],
    "retriever_stages": [
      "maxsim_rerank",
      "feature_search"
    ],
    "composes_with": [
      "feature_search (E5) -> maxsim_rerank on mixpeek://colbert_extractor@v1/colbertv2_v1"
    ]
  },
  "run": {
    "status": "served",
    "status_reason": "Ray Serve app in the inference registry; colbert_extractor batch uses ColBERTv2Batch. The embedding registry marks it PREVIEW.",
    "serve_apps": [
      "colbert_ir__colbertv2"
    ],
    "runtime": "transformers BertModel (add_pooling_layer=False) + AutoTokenizer, 768->128 linear head (ColBERTv2Encoder, engine code)",
    "weights_format": "safetensors",
    "trust_remote_code": false,
    "size_on_disk_mb": 438,
    "params_millions": 110,
    "preprocessing": "Documents: [CLS] [D] tokens [SEP], truncated to 512. Queries: [CLS] [Q] tokens [SEP], padded with [MASK] to 32 positions.",
    "postprocessing": "Project each token 768 -> 128 with the checkpoint linear head and L2-normalize; documents drop punctuation and padding tokens.",
    "serving_shape": "both",
    "batch_size": 16,
    "serve_batch": null,
    "scale_to_zero": {
      "min_replicas": 0,
      "max_replicas": 2,
      "idle_timeout_s": 300,
      "memory_gb": 2,
      "num_cpus": 1,
      "num_gpus": 0,
      "cold_start_s": {
        "warm_node_s": null,
        "new_node_s": null,
        "source": null
      },
      "cpu_fallback": true
    }
  },
  "cost": {
    "unmeasured": true,
    "measured": null,
    "note": "No committed throughput, latency, memory and cost measurement of the Serve app."
  },
  "evidence": {
    "unmeasured": false,
    "items": [
      {
        "dataset": "BEIR SciFact (5,183 documents)",
        "metric": "nDCG@10, exact MaxSim, engine encoder fp16 on MPS",
        "k": 10,
        "score": 0.7063,
        "baseline": null,
        "baseline_score": null,
        "held_out": false,
        "date": "2026-10-14",
        "source": "research/goal-specs/mvs-scale-from-zero/RESULTS-4.36-1M-RERUN.md"
      },
      {
        "dataset": "BEIR SciFact padded with 994,817 PubMed abstracts (1M documents)",
        "metric": "nDCG@10, exact MaxSim",
        "k": 10,
        "score": 0.4414,
        "baseline": null,
        "baseline_score": null,
        "held_out": false,
        "date": "2026-10-14",
        "source": "research/goal-specs/mvs-scale-from-zero/RESULTS-4.36-1M-RERUN.md"
      }
    ]
  },
  "refs": [
    "hf://colbert-ir/colbertv2.0@c1e84128e85ef755c096a95bdb06b47793b13acf"
  ],
  "formats": [
    {
      "ref": "hf://colbert-ir/colbertv2.0@c1e84128e85ef755c096a95bdb06b47793b13acf",
      "format": "safetensors",
      "file": "model.safetensors",
      "loader": "transformers / open_clip from_pretrained; Mixpeek Hub import (api/models/huggingface.py)",
      "supported_today": true
    },
    {
      "ref": "hf://colbert-ir/colbertv2.0@c1e84128e85ef755c096a95bdb06b47793b13acf",
      "format": "onnx",
      "file": "model.onnx",
      "loader": "onnxruntime; Mixpeek model upload (api/models/loadability.py, ONNX IR <= 11)",
      "supported_today": true
    },
    {
      "ref": "hf://colbert-ir/colbertv2.0@c1e84128e85ef755c096a95bdb06b47793b13acf",
      "format": "pytorch_bin",
      "file": "pytorch_model.bin",
      "loader": "torch.load (pickle); Mixpeek refuses pickle weights (shared/security/torch_pickle.py)",
      "supported_today": false
    }
  ],
  "use_cases": [
    "text-search",
    "reranking"
  ],
  "recommended": {
    "value": false,
    "for": [],
    "reason": null
  },
  "demos": [],
  "seo": {
    "title": "ColBERTv2: embedding model",
    "description": "ColBERTv2 is a text embedding model (MIT). Served on Mixpeek, scaling to zero when idle.",
    "keywords": [
      "ColBERTv2",
      "colbert",
      "embedding",
      "colbert-ir/colbertv2.0",
      "text-search",
      "reranking"
    ],
    "faqs": [
      {
        "q": "What is ColBERTv2 best at?",
        "a": "Late-interaction text retrieval and reranking (MaxSim over one 128-d vector per token); strong on out-of-domain and term-heavy queries."
      },
      {
        "q": "What are the limits of ColBERTv2?",
        "a": "Storage and cold reads: about 36 KB a 275-token document; English-centric BERT-base vocabulary; text only."
      },
      {
        "q": "Can ColBERTv2 be used commercially?",
        "a": "Yes. MIT on the Hugging Face repo; no use restrictions."
      },
      {
        "q": "How does ColBERTv2 run on Mixpeek?",
        "a": "Ray Serve app in the inference registry; colbert_extractor batch uses ColBERTv2Batch. The embedding registry marks it PREVIEW."
      }
    ]
  },
  "last_verified": "2026-11-05",
  "sources": [
    "https://huggingface.co/colbert-ir/colbertv2.0",
    "https://huggingface.co/api/models/colbert-ir/colbertv2.0",
    "https://arxiv.org/abs/2112.01488",
    "server/engine/inference/colbert_ir/colbertv2/encoder.py",
    "server/engine/inference/colbert_ir/colbertv2/services.py",
    "server/engine/extractors/colbert_extractor/v1/definition.py",
    "server/shared/retrievers/stages/sort/maxsim/models.py"
  ],
  "registry": {
    "colbert_ir__colbertv2": {
      "deployment": "colbert_ir__colbertv2",
      "service_id": "colbert-ir/colbertv2",
      "routes": "engine/inference/colbert_ir/colbertv2/routes.py",
      "route_prefix": "/colbert_ir/colbertv2/v1",
      "enabled": true,
      "description": "ColBERTv2 late-interaction text encoder (colbert-ir/colbertv2.0). Returns one 128-dimensional vector per token for MaxSim scoring: documents keep their non-punctuation tokens, queries are padded to 32.",
      "capabilities": {
        "category": "embedding",
        "modality_input": [
          "text"
        ],
        "modality_output": "vector",
        "supports_batch": true,
        "reranking_mode": null
      },
      "vector_config": {
        "dimensions": 128,
        "distance": "cosine",
        "vector_type": "multi_dense"
      },
      "serve": {
        "min_replicas": 0,
        "max_replicas": 2,
        "idle_timeout_s": 300,
        "target_ongoing_requests": 2,
        "num_cpus": 1,
        "num_gpus": 0,
        "memory_gb": 2,
        "pinnable": true,
        "placement": "worker",
        "enabled": true
      },
      "interactive_twin": false,
      "interactive_serve": null,
      "schemas": {
        "input": "shared.inference.colbert_ir.colbertv2.models.InferenceRequest",
        "output": "shared.inference.colbert_ir.colbertv2.models.InferenceResponse",
        "config": "shared.inference.colbert_ir.colbertv2.models.InferenceConfigs"
      }
    }
  },
  "templates": [],
  "usage": {
    "namespaces": null,
    "collections": null,
    "uncounted": true,
    "templates": 0,
    "counted_at": null,
    "source": null
  },
  "facets": {
    "status": "served",
    "category": "embedding",
    "family": "colbert",
    "license": "MIT",
    "commercial_use": true,
    "modalities_in": [
      "text"
    ],
    "modalities_out": [
      "multi_vector"
    ],
    "use_cases": [
      "text-search",
      "reranking"
    ],
    "recommended": false,
    "release_date": "2021-12",
    "params_millions": 110,
    "size_on_disk_mb": 438,
    "vector_dims": 128,
    "vector_distance": "cosine",
    "weights_format": "safetensors",
    "formats": [
      "onnx",
      "pytorch_bin",
      "safetensors"
    ],
    "loadable_today": true,
    "trust_remote_code": false,
    "cpu_fallback": true,
    "num_gpus": 0,
    "memory_gb": 2,
    "min_replicas": 0,
    "serving_shape": "both",
    "served": true,
    "ref_schemes": [
      "hf"
    ],
    "has_evidence": true,
    "cost_measured": false,
    "template_count": 0,
    "usage_namespaces": null,
    "usage_collections": null,
    "usage_uncounted": true,
    "has_demo": false,
    "rank_score": 4
  },
  "rank_score": 4
}