GenAI Features - Search AI

Back to Generative AI Features

LLM-powered features for Search AI that enable answer generation, vector search, document enrichment, and query processing.

The platform regularly integrates new models from providers like OpenAI, Azure OpenAI, and Anthropic. To use a model not yet available as a pre-built integration, add it using Provider’s New LLM Integration.

Model Feature Matrix

(✅ Supported | ❌ Not supported | ✅* Supported but no default prompt | NA = Not Applicable)

Answer Generation and Enrichment

For Enrich Chunks with LLM and Transform Documents with LLM, use templates from the prompt library to write custom prompts.

Model	Answer Generation	Enrich Chunks with LLM	Transform Documents with LLM
Azure OpenAI – GPT-4 Turbo	✅	✅*	✅*
Azure OpenAI – GPT-4o	✅	✅*	✅*
Azure OpenAI – GPT-4o mini	✅*	✅*	✅*
OpenAI – GPT-3.5 Turbo, GPT-4, GPT-4 Turbo	✅	✅*	✅*
OpenAI – GPT-4o	✅	✅*	✅*
OpenAI – GPT-4o mini	✅*	✅*	✅*
Custom LLM	✅*	✅*	✅*
XO GPT	✅	❌	❌
Amazon Bedrock	✅*	✅*	✅*

Query Processing

Model	Metadata Extractor Agent	Query Rephrase (Adv Search)	Query Transformation	Rephrase User Query	Result Type Classification
Azure OpenAI – GPT-4 Turbo	❌	❌	❌	✅*	❌
Azure OpenAI – GPT-4o	✅	✅	✅	✅	✅
Azure OpenAI – GPT-4o mini	✅*	✅*	✅*	✅*	✅*
OpenAI – GPT-3.5 Turbo, GPT-4, GPT-4 Turbo	❌	❌	❌	✅*	❌
OpenAI – GPT-4o	✅	✅	✅	✅	✅
OpenAI – GPT-4o mini	✅*	✅*	✅*	✅*	✅*
Custom LLM (GPT-4o / GPT-4o mini underlying)	✅*	✅*	✅*	✅*	✅*
XO GPT	❌	❌	❌	✅	❌
Amazon Bedrock	✅*	✅*	✅*	✅*	✅*

Vector Generation

Model	Vector Generation – Text	Vector Generation – Image
Azure OpenAI (all models)	NA	NA
OpenAI (all models)	NA	NA
Custom LLM	✅*	✅*
XO GPT	✅	✅
Amazon Bedrock	NA	NA

Supported vector dimensions for custom embedding model integrations:

128, 256, 384, 512, 768, 1024, 1028, 1536, 2048, 3072

See how to integrate a custom embedding model.

---

Features

Answer Generation

Generates an answer to the user’s question based on data ingested into the Search AI application. Relevant data is retrieved and inserted into the prompt; the configured LLM returns a formatted answer. Learn more.

Enrich Chunks with LLM

Uses an external LLM to refine, update, or enrich chunks extracted from ingested content. Learn more.

You must create a custom prompt to use this feature. All chunk fields are available for use in the prompt — click View Field Details when adding a Workbench Stage to see the full list.

Transform Documents with LLM

Uses an external LLM to enhance or update documents during the extraction process. Learn more.

You must create a custom prompt to use this feature. All document fields are available for use in the prompt — click View Field Details when adding a Transformation Stage to see the full list.

Vector Generation – Text

Creates vector embeddings for ingested text data. When a user submits a query, it is converted into an embedding and a vector search retrieves the most relevant data, which is then passed to answer generation.

Vector Generation – Image

Creates vector embeddings for ingested image data. When a user submits a query, it is converted into an embedding and a vector search retrieves the most relevant images, which are then passed to answer generation.

Metadata Extractor Agent

Extracts relevant sources and fields from a query, maps them to structured data, and applies filters or boosts for accurate retrieval. Particularly useful for data from third-party applications. Learn more.

If using a custom prompt, the LLM output must follow this structure:

{
  "extractedMetaData": [
    {
      "sourceName": "string",
      "chunkMeta": {
        "metaKey1": ["value1", "value2"],
        "metaKey2": ["value3"]
      }
    }
  ],
  "range": [
    {
      "createdOn": {
        "gte": "YYYY-MM-DDTHH:MM:SSZ",
        "lte": "YYYY-MM-DDTHH:MM:SSZ"
      }
    }
  ],
  "sourceIntent": true
}

• extractedMetaData: Array of sources and associated metadata.
• range: Optional date range filter.
• sourceIntent: Boolean indicating whether the source was explicitly specified.

Query Rephrase for Advanced Search API

Adds contextual information to user queries to enhance their relevance. Learn more.

If using a custom prompt, the LLM output must follow this structure:

{
  "rephrased_query": "string",
  "confidence": "High",
  "reasoning": "The rephrasing adds clarity and reflects the conversation context."
}

• rephrased_query: The reworded version of the original query.
• confidence: Confidence level in the quality of rephrasing.
• reasoning: Justification for the transformation.

Query Transformation

Identifies key terms within a query, removes noise, and prioritizes relevant documents. Learn more.

If using a custom prompt, the LLM output must follow this structure:

{
  "query_processing": {
    "original_query": "string",
    "keyword_search_query": "string",
    "vector_search_query": "string"
  },
  "core_terms": ["string"],
  "semantic_expansions": ["string"],
  "search_priority": {
    "must_include": ["string"],
    "should_include": ["string"],
    "context_terms": ["string"]
  }
}

Field	Description
query_processing.original_query	The input query provided by the user.
query_processing.keyword_search_query	Optimized for keyword-based search.
query_processing.vector_search_query	Adapted for vector-based semantic search.
core_terms	Key terms extracted from the query. Reserved for future use.
semantic_expansions	Related or semantically similar terms. Reserved for future use.
search_priority.must_include	Critical terms; strongest boost. Single-word terms allowed.
search_priority.should_include	Moderate relevance boost. Must contain at least two words.
search_priority.context_terms	Light contextual boosting. Must contain at least two words.

Example:

{
  "query_processing": {
    "original_query": "what are the large language models which have large context window in them",
    "keyword_search_query": "large language models large context window",
    "vector_search_query": "large language models which have large context window"
  },
  "search_priority": {
    "must_include": ["large language models", "large context window"],
    "should_include": ["models"],
    "context_terms": ["LLM"]
  }
}

Result Type Classification

Used in Agentic RAG to determine whether the user seeks a specific answer or a list of search results. Learn more.

If using a custom prompt, the LLM output must follow this structure:

{
  "query_type": "TYPE_1/TYPE_2",
  "confidence": "High",
  "reasoning": "This query closely matches the definition of TYPE_1 based on its intent and phrasing."
}

• query_type: TYPE_1 (Search Results) or TYPE_2 (Answers).
• confidence: Certainty level (for example, High, Medium, Low).
• reasoning: Brief explanation for the chosen type.

Rephrase User Query

Reconstructs incomplete or ambiguous user inputs using conversation history, improving intent detection and entity extraction accuracy. Handles three scenarios:

Scenario	Description	Example
Completeness	Completes an incomplete query using conversation context.	”How about Orlando?” → “What’s the weather forecast for Orlando tomorrow?”
Co-referencing	Resolves pronouns or vague references using prior context.	”Every six hours.” → “I take ibuprofen every six hours.”
Completeness + Co-referencing	Handles both issues together.	”What about interest rates?” → “What are the interest rates for personal and home loans?”

Conversation History Length

Controls how many previous messages are used as rephrasing context. Default: 5. Maximum: 25. Limited to the session’s available history.

Access from Rephrase User Query > Advanced Settings.

---