????[Public Preview] Understanding UiPath? Context Grounding: Enhancing GenAI Predictions with Your Business Data

UiPath Context Grounding is a crucial component of the UiPath AI Trust Layer, designed to optimize generative AI (GenAI) predictions by integrating your business data. This feature ensures that your data is prepared and ready for large language models (LLMs) without requiring additional subscriptions to embedding models, vector databases, or LLMs themselves. By creating representative indices and embeddings from your business data, UiPath GenAI features can access contextual evidence during runtime, improving the relevance and accuracy of AI-driven outputs.

Seamless Integration with UiPath GenAI

Context Grounding operates as a tenant-scoped platform service, supporting UiPath GenAI experiences—such as GenAI Activities—by embedding relevant information into user prompts before they are processed by LLMs via Retrieval-Augmented Generation (RAG).

?? Key Advantages of RAG Integration

The RAG service embedded within UiPath GenAI experiences provides several benefits:

• Overcoming Context Window Limitations: Whether dealing with small or large models, RAG enhances model accuracy, reliability, scalability, and efficiency by leveraging existing knowledge bases.
• Reducing Hallucination Risks: By referencing ground truth data, RAG minimizes the risk of generating inaccurate or misleading information.
• Access to Specialized Knowledge: Generative applications can tap into proprietary and specialized knowledge sources.
• Up-to-Date Information: Ensure that generative applications have access to the latest data and insights.
• Positive Feedback Loops: Create a dynamic interaction between data stores and user queries, enabling continuous improvement.

?? Core Components of Context Grounding

Context Grounding is composed of several key elements, each playing a vital role in preparing your business data for LLMs:

?? Ingestion and Indexing: Preparing Business Data

• Ingestion: Converts business data into embeddings using UiPath-managed models, making it LLM-ready.
• Embedding: Transforms business data into a format that LLMs can comprehend and search.
• Index: Organizes embeddings into a structured format within a vector database.
• Vector Databases: UiPath-managed databases that store and organize embeddings.

?? Semantic Similarity Search: Finding Relevant Information

• Search Functionality: Searches through LLM-ready business data to retrieve the most relevant information.
• Query Interpretation: Treats user prompts as queries to locate the best matches using cosine similarity search. This search is a precursor to RAG, augmenting prompts with relevant business data.

?? Retrieval-Augmented Generation (RAG)

• Prompt Grounding: Enriches and updates user prompts with the most relevant information identified in the semantic similarity search before executing the LLM generation through the AI Trust Layer's LLM Gateway.

?? UiPath GenAI Activities package

• UiPath GenAI Activities: Enables direct work with UiPath-managed large language models (LLMs) from third-party providers, integrated with UiPath AI Services and the AI Trust Layer for easy access to popular LLMs without managing subscriptions.
• Features: Includes the Content Generation activity for custom prompts and predefined activities like Summarize and Translate, all usable within Studio Web, Studio Desktop, and Apps.
• No Third-Party Subscription Required: Internal UiPath services handle request normalization and transformation, but connections need to be managed for workflows.
• Control and AI Units: GenAI Activities can be disabled via AI Trust Layer Automation Ops policy, with each activity execution consuming one AI Unit.

?? How Context Grounding interacts with your data in the GenAI Activities

Here's a summary of how Context Grounding interacts with your data in UiPath GenAI Activities:

• Data Source Establishment: Start by setting up a dataset (e.g., documents) in a shared Orchestrator folder, ensuring it's accessible for Context Grounding, which is tenant-scoped. You need Edit permissions to upload or manage files in this folder.
• Data Ingestion: Ingest your data into Context Grounding using activities like Index and Ingest (Public Preview) and Delete Index (Public Preview) to manage your data's lifecycle.
• Query and Prompt Grounding: Leverage the Content Generation activity to query over the ingested documents, using the retrieved information to augment or ground your prompts effectively.

?? STEP 1: Establish Connection with UiPath GenAI Activities

?? STEP 2 : Create Index Using "Index and Ingest" Activity (Public Preview)

The Index and Ingest (Public Preview) feature in UiPath GenAI Activities allows users to index and ingest data from various sources to create embeddings that enhance Retrieval Augmented Generation (RAG). This functionality is available on both Windows and cross-platform systems.

To configure the feature, users must select a connection ID from Integration Service, choose an Orchestrator folder and bucket for data storage, and specify the index name and data type (PDF, JSON, or CSV). The file glob pattern helps in defining the files to be ingested.

Output variables include Index ID, Datasource ID, and Index and Ingest results.

In practice, the activity makes datasets accessible for querying and RAG by large language models (LLMs). It involves creating indices in a vector database to store and reference embeddings and converting business data from Orchestrator buckets into these embeddings. Users can manage indices by creating new ones, re-ingesting data, or re-indexing existing datasets.

?? STEP 3 : Content Generation - (For Search/ Build FAQ/ Personal Assistant etc. ) Activity (Public Preview)

?? Result

Features of UiPath Context Grounding

UiPath Context Grounding offers a range of features to enhance your GenAI experiences:

• Multi-Document Support: Compatible with PDF, JSON, and CSV files, with plans to support more formats.
• Managed Ingestion and Indexing: UiPath optimizes data ingestion and indexing in its vector databases.
• Multiple Surfaces: Available through UiPath GenAI Activities.
• Semantic Similarity Search: Allows querying within documents or across datasets using advanced techniques to ensure highly relevant search results.
• RAG: Provides just-in-time grounding of prompts with the latest and most accurate information from business data.
• Proof of Knowledge: Generates citations from reference sources, offering transparency in AI-generated outputs.
• Streaming Support: Includes API support for real-time generation viewing.
• Data Source Integration: Seamlessly integrates with UiPath Orchestrator bucket entities, allowing data stored in shared folders to be ingested, indexed, and queried.
• Multilingual Support: Capable of handling documents in all UTF-8 encoded languages.

Limitations to Consider

• File Type Support: Currently limited to PDF, JSON, and CSV formats.
• Indexing Limits: Each tenant is limited to ten indices, and it is recommended to maintain a one-to-one relationship between indices and Orchestrator buckets.
• Version Compatibility: Requires UiPath Studio Web or Studio Desktop version 2024.4 or newer for full functionality.

References

1. Common Context Grounding patterns