Features¶

MRAG application comes with lot of features to enable a user to index the documents, chat with the index documents. Below are the details of the features present in MRAG application.

Provides Drag & Drop UI¶

You can build indexing pipelines using a drag-and-drop UI.
Simply drag and drop the components availabe on to the canvas and connect them to build a pipeline.

Provides Interactive QA Bot¶

You can chat with your documents by asking questions to the chatbot.
MRAG also supports voice chat where you can use your voice to ask a question.
The retrieved context will be shown on the left (relevant text will be highlighted in yellow) for the user to verify the correctness of the bot’s response.
Each indexing pipeline will have its own chatbot and associated settings.

Customizable QA Chat Bot Settings¶

MRAG allows a user to configure the following settings
- Prompt Template
- Query Settings
  
  Query Cleaning
  
  Query Rewriting
  
  Query Decomposition
  
  Self Querying
- Retrieval Techniques
  
  Simple Vector Retriever
  
  Multi Query / Query Expansion Retriever
  
  Reciprocal Rank Fusion (RRF) Retriever
- Retriever Settings
  
  Similarity Top K
  
  Sparse Top K
  
  Hybrid Top K
  
  Hybrid Search Alpha
- Reranker Settings
  
  Reranker Model
  
  Reranker Top K
- Context Compression / Denoising
  
  Context Extraction
  
  Context Filtering
- LLM Settings
  
  Model
  
  Model Temperature
- Post-processing of response
  
  Combine sub query responses
- Metrics
  
  Context Relevance Score
  
  Response Hallucination Score

Supports Query Enrichment¶

Enables users to enrich the query for better retrieval and LLM responses.
Supports query decomposition by generating sub queries to fetch answers for multiple queries in a single request.
Supports query cleaning and query rewriting for better retrieval and LLM response.

Enables Context Enrichment¶

Enables a user to enrich the context using HyPE technique.
HyPE technique generates hypothetical prompts / queries using an LLM that can be answered using the information in the chunk.
User can select the number of hypothetical queries to generate.
Improves retrieval quality by comparing user’s query with hypothetical queries for similarity.

Provides Evaluation Metrics¶

Enables users to evaluate the performance of the retriever and LLM responses.
Context relevance score helps in evaluating the retriever’s performance.
Response hallucination score helps in evaluation the LLM response quality.

Provides Query Suggestions¶

Provides query suggestions to the user if a user includes HyPE in the indexing pipeline.
Enables user to view the query suggestions for each document separately.
Helps a user in getting an overview of the questions a document can answer.

Provides Multiple Splitters¶

Provides 5 splitters to chunk the documents.
Supports custom splitters like Regex Splitter, PDF Font Splitter and Dummy Splitter.
User can apply different splitter for different documents based on the document structure which ensures smart chunking.

Provides Metadata Extractor¶

Provides metadata extractor to extract metadata using regular expressions from documents.
Extracted metadata can be used in splitters to enable self querying.
Self querying enables user to filter and retrieve the documents based on metadata.

Enables Metadata Testing¶

Enables a user to test the metadata extraction by uploading a sample document.
The user can view the metadata extracted using the schema defined by the user.
The user can also see the raw text of the document along with the extracted metadata.

Provides Document Viewer¶

Enables a user to view the raw file content as well as the fonts in a PDF file.
This enables the user to decide the chinking strategy to use.
Aids in smart chunking strategy as per the document structure.

Provides Document Chunk Viewer¶

Enables users to view the chunks before building an indexing pipeline.
Enables users to compare multiple splitters and choose the best one for their documents.
Enables users to apply multiple splitters to a document.

Supports Text and PDF Documents¶

Supports text and pdf document types.
Users can upload the documents from their local storage.
Total size of documents should be less than 5MB.
Individual document size must be less than 1MB.