Skip to main content
The RAG framework is built on a set of core principles designed to deliver flexible, context-driven AI solutions across applications. Each of the following core concepts is a key component of the framework.

Core Concepts

Applications

Each application using the RAG framework has its own configuration, tailored to its unique requirements. Applications can define specific system prompts, custom tools, enabled agents, and LLM model preferences to align with their individual workflows and data needs.

Tool Repository

The Tool Repository is a centralized collection of tools used by agents to complete tasks. It includes general-purpose tools, which offer broad functionality, and app-specific tools, which are custom-built to meet unique application requirements. Each tool follows a structured input and output schema for consistent integration.

Agents

Agents are specialized AI entities configured with profiles that determine their behavior, workflow type, and the tools they can access. Agents process user requests by selecting and sequencing tools.

Dynamically Generated Pipelines

Agents generate pipelines by breaking down complex tasks into subtasks, each associated with a specific tool. These pipelines follow an Input-Process-Output (IPO) loop with error-handling and self-correction capabilities, allowing the agent to adapt dynamically and handle complex workflows.

Vectorization

Vectorization is the process of transforming content into vector embeddings to enable efficient and semantically relevant retrieval.

Capabilities

NameDescription
Agentic WorkflowsEnables autonomous workflows where AI agents make independent decisions based on task context.
Agent Intent AnalysisAligns agent choices with user intent.
Dynamic Pipeline GenerationConstructs task pipelines from available tools, selecting the best tools for precise processing.
Iterative Self-CorrectionRefines responses through iterative feedback, dynamically handling errors to improve accuracy.
Knowledge Base RetrievalUses a knowledge base for accurate, content-driven answers requiring information retrieval.
Automated Document HandlingAutomates document processing and categorization.
Contextual RetrievalRetrieves data based on surrounding context, ensuring highly relevant, contextually situated responses.
Hypothetical Document Embeddings (HyDE)Enhances embeddings by generating hypothetical answers, improving understanding of complex queries.
Hybrid SearchCombines vector similarity with keyword-based search, ensuring relevance across varied query types.
Real-Time Re-rankingRe-orders search results by contextual relevance, prioritizing the most pertinent information.
Date-Based Vector CalculationPrefers the most recent documents during re-ranking, prioritizing up-to-date information for time-sensitive queries.
Multi-Source VectorizationIngests and vectorizes data from PDFs, RSS feeds, and structured databases, enhancing retrieval capabilities.
Data-Based RetrievalRetrieves structured data from specific databases, allowing precise, data-driven responses.
External Data Source IntegrationSupports real-time data from external sources like APIs or RSS feeds, keeping responses relevant.
Data Aggregation & SummarizationSummarizes large datasets, providing easy-to-interpret information at a glance.
Structured OutputProvides schema-based outputs aligned with templates, ideal for reports and standardized responses.
Visualization CapabilitiesSupports charts, tables, and graphs for visual representation of complex data.
Proactive Query SuggestionsSuggests additional questions based on conversation flow, guiding interactions for deeper insights.
User-Specific Context RetentionRetains user-specific context across sessions for personalized experiences.
Embeddings ManagementManages vector embeddings for efficient similarity searches and relevant content retrieval.
Guards (Safe Input/Output)Implements safeguards for data handling, validating inputs and outputs for security and integrity.
Automated AI TestingContinuous testing of AI models to ensure consistent, secure, and reliable performance across scenarios.
Fine-Tuning and OptimizationAdapts pre-trained models for specific tasks, improving relevance and performance for specialized needs.
Memory-Driven CachingCaches responses based on user intention, reducing redundancy and enhancing response time.
Tool Argument-Based CachingCaches results based on tool arguments, optimizing responses for repeated queries.
SDK IntegrationProvides a JavaScript and React SDK, enabling easy frontend integration and customization.
Adaptive NLP ModelsUses tailored NLP models to suit different conversational contexts.