RAG Demo using Couchbase, Streamlit, Haystack, and OpenAI

This is a demo app built to chat with your custom PDFs using Couchbase Vector Search to augment the OpenAI results in a Retrieval-Augmented-Generation (RAG) model.

Three Implementation Options

Option 1: Search Vector Index (chat_with_pdf_with_search_vector_index.py)

Uses CouchbaseSearchDocumentStore with Search vector indexes, which offers:

• Flexible vector search
• Rich text search combined with vector similarity
• Complex filtering using FTS queries
• Compatible with Couchbase 7.6+
• Ideal for hybrid search scenarios combining full-text and vector search

Option 2: Hyperscale Vector Index (Default - chat_with_pdf_with_query_vector_index.py)

Uses CouchbaseQueryDocumentStore with Hyperscale vector index, which offers:

• High-performance vector search at massive scale (billions of documents)
• Pure vector search optimized for RAG applications
• SQL++ queries for efficient vector retrieval
• Recommended for Couchbase 8.0+ for pure vector similarity search

Option 3: Composite Vector Index

Uses CouchbaseQueryDocumentStore with Composite vector index, which offers:

• Vector search with metadata filtering
• Combines vector fields with scalar fields for pre-filtering
• SQL++ queries with efficient filtered vector retrieval
• Best for filtered vector search scenarios (e.g., filter by date, category, user_id)
• Recommended for Couchbase 8.0+ when you need to filter before vector search

This demo doesn't use Composite Vector index, but you can easily do so by just removing VECTOR from this line and keeping the rest same. To learn more about how Composite Vector Indexes are made, you can refer here.

How does it work?

You can upload your PDFs with custom data & ask questions about the data in the chat box. For each question, you will get two answers:

• one using RAG (Couchbase logo)
• one using pure LLM - OpenAI (🤖).

The RAG pipeline utilizes Haystack, Couchbase Vector Search, and OpenAI models. It fetches relevant parts of the PDF using vector search & adds them as context for the language model.

Quick Start

1. Clone this repository

   git clone <repository-url>
   cd haystack-demo

2. Create a Python virtual environment

   python -m venv venv
   source venv/bin/activate

3. Install dependencies

   pip install -r requirements.txt

4. Create a Couchbase bucket (via Couchbase UI/Capella) with the name "sample_bucket"

5. Configure environment variables (see Setup section below)

6. Run the Streamlit app

   # For Hyperscale Vector Index (default)
   streamlit run chat_with_pdf_with_query_vector_index.py
   
   # OR for Search Vector Index
   streamlit run chat_with_pdf_with_search_vector_index.py

7. Upload a PDF - everything else is automatic!

The app automatically creates:

• Scopes and collections
• Vector indexes (after PDF upload for chat_with_pdf_with_query_vector_index.py, or on startup for chat_with_pdf_with_search_vector_index.py)

Which Option Should You Choose?

Couchbase Capella supports three types of vector indexes:

• Hyperscale Vector Index (chat_with_pdf_with_query_vector_index.py) - Best for RAG/chatbot applications with pure semantic search and billions of documents
• Composite Vector Index - Best when you need to filter by metadata before vector search
• Search Vector Index (chat_with_pdf_with_search_vector_index.py) - Best for hybrid searches combining keywords, geospatial, and semantic search

For this PDF chat demo, we recommend Hyperscale Vector Index for optimal performance in RAG applications.

Learn more about choosing the right vector index in the official Couchbase vector index documentation.

Setup and Installation

Install dependencies

pip install -r requirements.txt

Set the environment secrets

Copy the secrets.example.toml file in .streamlit folder and rename it to secrets.toml and replace the placeholders with the actual values for your environment

For Hyperscale Vector Index (chat_with_pdf_with_query_vector_index.py):

DB_CONN_STR = "<couchbase_cluster_connection_string>"
DB_USERNAME = "<couchbase_username>"
DB_PASSWORD = "<couchbase_password>"
DB_BUCKET = "<bucket_name>"
DB_SCOPE = "<scope_name>"
DB_COLLECTION = "<collection_name>"
OPENAI_API_KEY = "<openai_api_key>"

For Search Vector Index (chat_with_pdf_with_search_vector_index.py):

Add one additional environment variable to the above configuration:

INDEX_NAME = "<vector_capable_search_vector_index_name>"

Automatic Resource Setup

The application automatically handles resource creation in the following order:

On Application Startup:

1. Creates the scope if it doesn't exist
2. Creates the collection if it doesn't exist

After PDF Upload (chat_with_pdf_with_query_vector_index.py):

3. Automatically creates the Hyperscale index after documents are loaded
4. Falls back to creating the index on first query if needed

On Application Startup (chat_with_pdf_with_search_vector_index.py):

3. Attempts to create the Search Vector index (can be created without documents)

What You Need:

• Your Couchbase bucket must exist with the name "sample_bucket"
• All other resources (scope, collection, indexes) are created automatically
• No manual index creation required - just upload your PDF and the index will be created

Note: For chat_with_pdf_with_query_vector_index.py, the vector index is created automatically after you upload your first PDF because Hyperscale/Composite indexes require documents for training.

Manual Vector Index Creation (Optional)

⚠️ Manual creation is NOT required - the app creates indexes automatically when you upload a PDF. This section is only for advanced users who want manual control.

Hyperscale or Composite Vector Index

You need to create a Hyperscale or Composite vector index on your collection after loading some documents (required for index training). Choose between Hyperscale or Composite Index based on your use case. Whichever vector index (Hyperscale or Composite) you choose won't affect the functionality of this demo, though performance differences may occur.

Option 1: Hyperscale Vector Index (Recommended)

Hyperscale is a dedicated vector index optimized for pure vector search at massive scale. Use this for the best performance in RAG applications. Refer to the Hyperscale Vector Index Guide for detailed instructions.

Creating a Hyperscale Index using SQL++ (use Couchbase Query Workbench or programmatically):

CREATE VECTOR INDEX idx_pdf_hyperscale
ON `bucket_name`.`scope_name`.`collection_name`(embedding VECTOR) 
WITH {
  "dimension": 1536,
  "similarity": "DOT"
};

Option 2: Composite Vector Index

Composite indexes combine vector fields with other scalar fields. This is useful when you need to filter documents by metadata before performing vector search.

Creating a Composite Index using SQL++:

CREATE INDEX idx_pdf_composite 
ON `bucket_name`.`scope_name`.`collection_name`(embedding VECTOR) 
WITH {
  "dimension": 1536,
  "similarity": "DOT"
};

Index Parameters:

• dimension: Must match your embedding model (1536 for OpenAI ada-002/ada-003, 768 for sentence-transformers)
• similarity: Must match the similarity metric in CouchbaseQueryDocumentStore. Use DOT for dot product (recommended for OpenAI embeddings)

Important Notes:

1. Index Creation Timing: Hyperscale and Composite vector indexes require training data. Create the index after you've loaded the documents into your collection.
2. Similarity Metric: The similarity parameter in the index must match the similarity parameter in your CouchbaseQueryDocumentStore configuration.
3. Dimension: Must match your embedding model's output dimensions.

Verifying Your Index:

After creating the index, verify it exists:

SELECT * FROM system:indexes 
WHERE name LIKE 'idx_%_vector';

Search Vector Index (for chat_with_pdf_with_search_vector_index.py)

Automatic Creation: The app attempts to create the Search Vector index automatically on startup using the INDEX_NAME from your configuration.

Manual Creation (if automatic creation fails): Create a Full Text Search index with vector capabilities.

Creating a Search Vector Index with Vector Support

If automatic creation fails, you can create the index using the Couchbase UI or by importing the provided index definition.

Using Couchbase Capella:

1. Follow the import instructions here
2. Use the provided sampleSearchIndex.json file in this repository
3. Update the following values in the JSON before importing:
   • sourceName: Replace sample_bucket with your bucket name
   • types: Replace scope.coll with your actual scope_name.collection_name
4. Import the file in Capella
5. Click on Create Index

Using Couchbase Server:

1. Navigate to Search -> Add Index -> Import
2. Use the provided sampleSearchIndex.json file in this repository
3. Update the following values in the JSON before importing:
   • sourceName: Replace sample_bucket with your bucket name
   • types: Replace scope.coll with your actual scope_name.collection_name
4. Paste the updated JSON in the Import screen
5. Click on Create Index

Search Vector Index Definition

The sampleSearchIndex.json file contains a pre-configured Search Vector index with vector capabilities. Key features:

• Index Name: sample-index (customizable)
• Vector Field: embedding with 1536 dimensions
• Similarity: dot_product (optimized for OpenAI embeddings)
• Text Field: content for document text
• Metadata: Dynamic mapping for meta fields

Run the Application

For Hyperscale or Composite Vector Index:

streamlit run chat_with_pdf_with_query_vector_index.py

For Search Vector Index:

streamlit run chat_with_pdf_with_search_vector_index.py

Implementation Details

Hyperscale Vector Index Implementation (chat_with_pdf_with_query_vector_index.py)

This demo uses the following key components:

1. CouchbaseQueryDocumentStore:

   • Configured with QueryVectorSearchType.ANN for fast approximate nearest neighbor search
   • Uses QueryVectorSearchSimilarity.DOT for dot product similarity (recommended for OpenAI embeddings)
   • Supports both Hyperscale and Composite vector indexes
   • Leverages SQL++ for efficient vector retrieval
   • Same code works for both index types - just create the appropriate index

2. CouchbaseQueryEmbeddingRetriever:

   • Uses SQL++ queries with APPROX_VECTOR_DISTANCE() function for ANN search
   • Retrieves top-k most similar documents based on embedding similarity
   • Optimized for low-latency, high-throughput vector search

3. OpenAI Embeddings:

   • text-embedding-ada-002 model with 1536 dimensions
   • Generates embeddings for both documents and queries

For more details on implementation, refer to the extensive code comments in chat_with_pdf_with_query_vector_index.py.

Search Vector Index Implementation (chat_with_pdf_with_search_vector_index.py)

This alternative implementation uses:

1. CouchbaseSearchDocumentStore:

   • Uses Full Text Search service for vector indexing and retrieval
   • Compatible with Couchbase 7.6+ and 8.0+
   • Supports rich text search combined with vector similarity

2. CouchbaseSearchEmbeddingRetriever:

   • Leverages Search vector index capabilities
   • Retrieves top-k most similar documents using FTS queries
   • Supports complex filtering with FTS query syntax

3. OpenAI Embeddings:

   • Same text-embedding-ada-002 model with 1536 dimensions
   • Generates embeddings for both documents and queries

For more details on FTS implementation, refer to the code comments in chat_with_pdf_with_search_vector_index.py.

Additional Resources

• Couchbase Vector Index Documentation
• Haystack Documentation
• couchbase-haystack GitHub Repository