Mockingbird: Vectara's LLM

Generating accurate and reliable summaries from large datasets can be a challenge. Mockingbird is a cutting-edge Large Language Model (LLM) developed by Vectara specifically for Retrieval Augmented Generation (RAG) use cases. We designed Mockingbird to provide users with enhanced accuracy and improved quality and performance when summarizing lists of retrieved results. Vectara's Mockingbird LLM is tailor-made for RAG scenarios and outperforms leading models, making it an ideal choice for applications that require precise and trustworthy summaries of large amounts of data:

• Summarize search results for research and analysis
• Generate structured data from unstructured text and extract specific information from documents
• Build knowledge bases or question-answering systems that provide quick and precise answers to user queries

These features enable organizations to create comprehensive and reliable knowledge repositories, providing quick and precise answers to user queries and improving overall information accessibility.

Significantly improved quality for Retrieval Augmented Generation (RAG)

Mockingbird improves quality for RAG use cases, surpassing general-purpose LLMs in critical areas for enterprise applications. Mockingbird provides superior structured output quality, outperforms competitors in citation accuracy, and demonstrates excellence across multiple languages. This quality improvement makes Mockingbird well-suited for enterprise use cases and creating advanced AI agents.

Increased accuracy in summarizing retrieved results

Mockingbird excels at producing summaries and answers for large collections of search results, allowing users to quickly grasp essential information without manual review. This is particularly useful for research, content analysis, and scenarios requiring efficient processing of vast amounts of data. Mockingbird meets or exceeds leading models like GPT-4 and Gemini 1.5 Pro in RAG quality, citation quality, multilingual quality and structured output accuracy, especially being a smaller and lower cost model, while reducing hallucinations and providing more reliable and accurate summaries for important data

Enhanced JSON output for structured data generation

Mockingbird effectively generates structured data from unstructured text sources. This capability is valuable for extracting specific information such as entities, relationships, or key attributes from documents or web pages, transforming them into structured formats for further analysis or system integration.

Multilingual capabilities

Mockingbird supports the following languages: Arabic, French, Spanish, Portuguese, Italian, German, Chinese, Dutch, Korean, Japanese, and Russian.

caution

Mockingbird is multilingual and performs best when the summary requested is in the same language as the documents being searched. Cross-lingual capabilities are not officially supported yet, but may work in some cases such as referencing documents in language X, and a summary in language Y.

Secure deployment within Vectara's infrastructure

Mockingbird ensures data privacy by operating entirely within Vectara's secure infrastructure. Unlike some providers who face accusations of training on customer data, Vectara guarantees your data is never used to train or improve our models, ensuring data privacy and compliance with the strictest security standards.

Selecting Mockingbird for summarization

To use Mockingbird in the Vectara Console:

1. Select Corpora from the main menu and go to a corpus.
2. Select the Query tab.
3. Click Model from the Generation drop-down in the Corpus Query Configuration panel.
4. Select the Mockingbird model.
5. Click Model again to minimize the list of models. This example shows a Summary configuration with Mockingbird selected as the model.

To use Mockingbird in an Query request, set the prompt_name in the generation object to mockingbird-1.0-2024-07-16:

{
"query:" "What is the infinite probability drive?",
"generation": {
    "prompt_name": "mockingbird-1.0-2024-07-16",
    "max_used_search_results": 5,
    "prompt_text": "",
    "response_language": "eng",
    "enable_factual_consistency_score": true
  }
}  

tip

Check out our interactive API reference that lets you run queries directly from your browser.

Default Mockingbird prompt

Mockingbird uses the following prompt_text by default. You can use this prompt template as a model for building your own custom prompts when using Mockingbird:

{
  "role": "user", "content": "You are a search bot that takes search results and 
  summarizes them as a coherent answer. Only use information provided in this chat. 
  Generate a comprehensive and informative answer for the query \n\n <query>" ${vectaraQuery} "</query> \n\n 
  solely based on following search results:\n\n 
  #foreach ($qResult in $vectaraQueryResults) \n [$foreach.index + 1) 
  ${qResult.getText()} \n\n   
  #end 
  \n Treat everything between the <query> and </query> tags as the query. 
  You must only use information from the provided results. Combine search results 
  together into a coherent answer. Do not repeat text. Cite search results using 
  [number] notation. Only cite the most relevant results that answer the question 
  accurately. If the search results are not valid, respond with - No result found. 
  Please generate your answer in the language of $vectaraLangName"
}

Custom prompts and prompt templates using Mockingbird have the following rules:

1. You are only allowed to specify the assistant and user roles.
2. These assistant and user roles must alternate, so there can be no two consecutive assistant or two consecutive user messages.
3. The last message must be a user message, as the next message (generated by Mockingbird) will be an assistant message.