Vectara
Log inBook a demo

Vectara's models

Vectara's advanced end-to-end agentic models have pioneered enterprise-grade agent context and accuracy.

Try our models

Best-of-breed models at every stage

Agents are only as good as the models that power them, so we maintain state-of-the-art models for every step in the process.

Vectara modelYour inputs and infrastructure
  1. Query

    A question from a person or an agent

  2. BoomerangRetrieval model
    • Stateless hybrid search
    • Coarse-grained sorting of results
  3. SlingshotReranker model

    Fine-grained sorting of results

  4. MockingbirdGenerative LLM

    Knows how to form accurate answers based on results

  5. HHEMHallucination evaluation

    Detects hallucinations by checking the response against the results

  6. Your data

    Expresses meaning in words (tokens)

  7. Vector database

    Stores meaning in vectors

  8. Response + factual consistency score

Boomerang v2

Vectara’s advanced retrieval and embedding model delivers high-precision, contextually relevant search results.

Documentation

Boomerang’s performance benchmarks

Performance details (September 2026)

Semiconductor Retrieval Benchmark

Boomerang v1 is text-only. Amazon Nova 2 MM can't embed a figure with its caption.

Semiconductor Retrieval Benchmark
ModelText, nDCG@10Image only, nDCG@5Image with captions, nDCG@5
Boomerang v20.6150.6120.758
Boomerang v10.520Text-onlyText-only
Gemini Embedding 20.5860.7100.662
Qwen3-VL-Embedding-2B0.4240.5030.679
Amazon Nova 2 MM0.5580.535Text/image-only

Slingshot

Vectara’s reranking model reorganizes search results to prioritize highly relevant and useful information.

Documentation

Slingshot’s performance benchmarks

Performance details (May 2024)

Vectara rerank vs other models, by domain

Higher is better

Vectara rerank vs other models, by domain
ModelDevShoppingNewsWikiSocial/Forum
Vectara Rerank0.850.450.570.850.79
Cohere Rerank V30.840.410.550.840.79
BCE-reranker-base0.820.400.510.820.76
Mono MT50.810.400.480.810.74
BGE-reranker-base0.810.370.450.810.75

Mockingbird v3 MM

Vectara’s multi-modal generative model is fine-tuned for semiconductor tasks.

Mockingbird’s performance benchmarks

Development process (October 2026)

Semiconductor Benchmarks

MMCircuitEval: public multimodal circuit benchmark

Semiconductor Benchmarks
ModelInternal Semiconductor BenchmarkMMCircuitEval
Mockingbird V3 MM0.65730.4763
GPT-6 Astra0.65450.4796
Inkling-Small0.69220.4734
Gemma4 31B Base0.57810.4572

Hughes Hallucination Evaluation Model (HHEM)

Vectara’s HHEM model is the most production-ready generative AI hallucination detector

Documentation

Our Hallucination Evaluation Model is a first-of-its-kind initiative to proffer a commercially available and open-source model that addresses the accuracy and level of hallucination in LLMs.

Simon Hughes (1979 – 2023) avatar
Simon Hughes (1979 – 2023)ML Engineer & AI Researcher Vectara

HHEM’s hallucination detection performance benchmarks

Performance details (August 2024)

Hallucination detection, F1 score

Higher is better

Hallucination detection, F1 score
ModelAggreFact-SOTARAGTruth-QARAGTruth-Summarization
HHEM v268.6%67.1%55.6%
GPT-463.9%57.8%40.6%
GPT-3.5 Turbo60.9%25.0%29.7%

Remembering a champion of responsible AI

The Hughes Hallucination Evaluation Model honors Dr. Simon Mark Hughes, a pioneering machine learning engineer at Vectara. His research on AI hallucination detection, before his passing in 2023, continues to shape industry efforts toward truthful and reliable AI outputs.

Create AI Agents in minutes

Vectara is the Agent OS for trusted enterprise AI: governed, grounded, auditable agents with real-time policy and factual-consistency enforcement.

Get started