Merge pull request #296 from ajithvcoder/feature_rag_tutorial

Add rag basic tutuorials

docs/source/use_cases/rag_documents.rst

@@ -0,0 +1,322 @@
+.. raw:: html
+
+   <div style="display: flex; justify-content: flex-start; align-items: center; margin-bottom: 20px;">
+      <a href="https://colab.research.google.com/github/SylphAI-Inc/AdalFlow/blob/main/notebooks/tutorials/adalflow_rag_documents.ipynb" target="_blank" style="margin-right: 20px;">
+         <img alt="Open In Colab" src="https://colab.research.google.com/assets/colab-badge.svg" style="height: 20px;">
+      </a>
+
+      <a href="https://github.com/SylphAI-Inc/AdalFlow/tree/main/tutorials/adalflow_rag_documents.py" target="_blank" style="display: flex; align-items: center;">
+         <img src="https://github.githubassets.com/images/modules/logos_page/GitHub-Mark.png" alt="GitHub" style="height: 20px; width: 20px; margin-right: 5px;">
+         <span style="vertical-align: middle;"> Open Source Code</span>
+      </a>
+   </div>
+
+RAG for documents
+=============================
+
+Overview
+--------
+
+This implementation showcases an end-to-end RAG system capable of handling large-scale text files and
+generating context-aware responses. It is both modular and extensible, making it adaptable to various
+use cases and LLM APIs.
+
+**Imports**
+
+- **SentenceTransformer**: Used for creating dense vector embeddings for textual data.
+- **FAISS**: Provides efficient similarity search using vector indexing.
+- **tiktoken**: ensures that the text preprocessing aligns with the tokenization requirements of the underlying language models, making the pipeline robust and efficient.
+- **GroqAPIClient and OpenAIClient**: Custom classes for interacting with different LLM providers.
+- **ModelType**: Enum for specifying the model type.
+
+.. code-block:: python
+
+    import os
+    import tiktoken
+    from typing import List, Dict, Tuple
+    import numpy as np
+    from sentence_transformers import SentenceTransformer
+    from faiss import IndexFlatL2
+
+    from adalflow.components.model_client import GroqAPIClient, OpenAIClient
+    from adalflow.core.types import ModelType
+    from adalflow.utils import setup_env
+
+The ``AdalflowRAGPipeline`` class sets up the Retrieval-Augmented Generation (RAG) pipeline. Its ``__init__`` method initializes key components:
+
+- An embedding model (``all-MiniLM-L6-v2``) is loaded using ``SentenceTransformer`` to convert text into dense vector embeddings with a dimensionality of 384.
+- A FAISS index (``IndexFlatL2``) is created for similarity-based document retrieval.
+- Parameters such as ``top_k_retrieval`` (number of documents to retrieve) and ``max_context_tokens`` (limit on token count in the context) are configured.
+- A tokenizer (``tiktoken``) ensures precise token counting, crucial for handling large language models (LLMs).
+
+The method also initializes storage for documents, their embeddings, and associated metadata for efficient management and retrieval.
+
+The ``AdalflowRAGPipeline`` class provides a flexible pipeline for Retrieval-Augmented Generation (RAG),
+initializing with parameters such as the embedding model (``all-MiniLM-L6-v2`` by default), vector dimension,
+top-k retrieval count, and token limits for context. It utilizes a tokenizer for token counting, a
+SentenceTransformer for embeddings, and a FAISS index for similarity searches, while also maintaining
+document data and metadata. The ``load_text_file`` method processes large text files into manageable chunks
+by splitting the content into fixed line groups, facilitating easier embedding and storage. To handle
+multiple files, ``add_documents_from_directory`` iterates over text files in a directory, embeds the content,
+and stores them in the FAISS index along with metadata. Token counting is achieved via the ``count_tokens``
+method, leveraging a tokenizer to precisely determine the number of tokens in a given text. The
+``retrieve_and_truncate_context`` method fetches the most relevant documents from the FAISS index based on
+query embeddings, truncating the context to adhere to token limits. Finally, the ``generate_response`` method
+constructs a comprehensive prompt by combining the retrieved context and query, invokes the provided model
+client for a response, and parses the results into a readable format. This pipeline demonstrates seamless
+integration of text retrieval and generation to handle large-scale document queries effectively.
+
+
+.. code-block:: python
+
+    class AdalflowRAGPipeline:
+        def __init__(self,
+                    model_client=None,
+                    model_kwargs=None,
+                    embedding_model='all-MiniLM-L6-v2',
+                    vector_dim=384,
+                    top_k_retrieval=3,
+                    max_context_tokens=800):
+            """
+            Initialize RAG Pipeline for handling large text files
+
+            Args:
+                embedding_model (str): Sentence transformer model for embeddings
+                vector_dim (int): Dimension of embedding vectors
+                top_k_retrieval (int): Number of documents to retrieve
+                max_context_tokens (int): Maximum tokens to send to LLM
+            """
+            # Initialize model client for generation
+            self.model_client = model_client
+
+            # Initialize tokenizer for precise token counting
+            self.tokenizer = tiktoken.get_encoding("cl100k_base")
+
+            # Initialize embedding model
+            self.embedding_model = SentenceTransformer(embedding_model)
+
+            # Initialize FAISS index for vector similarity search
+            self.index = IndexFlatL2(vector_dim)
+
+            # Store document texts, embeddings, and metadata
+            self.documents = []
+            self.document_embeddings = []
+            self.document_metadata = []
+
+            # Retrieval and context management parameters
+            self.top_k_retrieval = top_k_retrieval
+            self.max_context_tokens = max_context_tokens
+
+            # Model generation parameters
+            self.model_kwargs = model_kwargs
+
+        def load_text_file(self, file_path: str) -> List[str]:
+            """
+            Load a large text file and split into manageable chunks
+
+            Args:
+                file_path (str): Path to the text file
+
+            Returns:
+                List[str]: List of document chunks
+            """
+            with open(file_path, 'r', encoding='utf-8') as file:
+                # Read entire file
+                content = file.read()
+
+            # Split content into chunks (e.g., 10 lines per chunk)
+            lines = content.split('\n')
+            chunks = []
+            chunk_size = 10  # Adjust based on your file structure
+
+            for i in range(0, len(lines), chunk_size):
+                chunk = '\n'.join(lines[i:i+chunk_size])
+                chunks.append(chunk)
+
+            return chunks
+
+        def add_documents_from_directory(self, directory_path: str):
+            """
+            Add documents from all text files in a directory
+
+            Args:
+                directory_path (str): Path to directory containing text files
+            """
+            for filename in os.listdir(directory_path):
+                if filename.endswith('.txt'):
+                    file_path = os.path.join(directory_path, filename)
+                    document_chunks = self.load_text_file(file_path)
+
+                    for chunk in document_chunks:
+                        # Embed document chunk
+                        embedding = self.embedding_model.encode(chunk)
+
+                        # Add to index and document store
+                        self.index.add(np.array([embedding]))
+                        self.documents.append(chunk)
+                        self.document_embeddings.append(embedding)
+                        self.document_metadata.append({
+                            'filename': filename,
+                            'chunk_index': len(self.document_metadata)
+                        })
+
+        def count_tokens(self, text: str) -> int:
+            """
+            Count tokens in a given text
+
+            Args:
+                text (str): Input text
+
+            Returns:
+                int: Number of tokens
+            """
+            return len(self.tokenizer.encode(text))
+
+        def retrieve_and_truncate_context(self, query: str) -> str:
+            """
+            Retrieve relevant documents and truncate to fit token limit
+
+            Args:
+                query (str): Input query
+
+            Returns:
+                str: Concatenated context within token limit
+            """
+            # Retrieve relevant documents
+            query_embedding = self.embedding_model.encode(query)
+            distances, indices = self.index.search(
+                np.array([query_embedding]),
+                self.top_k_retrieval
+            )
+
+            # Collect and truncate context
+            context = []
+            current_tokens = 0
+
+            for idx in indices[0]:
+                doc = self.documents[idx]
+                doc_tokens = self.count_tokens(doc)
+
+                # Check if adding this document would exceed token limit
+                if current_tokens + doc_tokens <= self.max_context_tokens:
+                    context.append(doc)
+                    current_tokens += doc_tokens
+                else:
+                    break
+
+            return "\n\n".join(context)
+
+        def generate_response(self, query: str) -> str:
+            """
+            Generate a response using retrieval-augmented generation
+
+            Args:
+                query (str): User's input query
+
+            Returns:
+                str: Generated response incorporating retrieved context
+            """
+            # Retrieve and truncate context
+            retrieved_context = self.retrieve_and_truncate_context(query)
+
+            # Construct context-aware prompt
+            full_prompt = f"""
+            Context Documents:
+            {retrieved_context}
+
+            Query: {query}
+
+            Generate a comprehensive response that:
+            1. Directly answers the query
+            2. Incorporates relevant information from the context documents
+            3. Provides clear and concise information
+            """
+
+            # Prepare API arguments
+            api_kwargs = self.model_client.convert_inputs_to_api_kwargs(
+                input=full_prompt,
+                model_kwargs=self.model_kwargs,
+                model_type=ModelType.LLM
+            )
+
+            # Call API and parse response
+            response = self.model_client.call(
+                api_kwargs=api_kwargs,
+                model_type=ModelType.LLM
+            )
+            response_text = self.model_client.parse_chat_completion(response)
+
+            return response_text
+
+The ``run_rag_pipeline`` function demonstrates how to use the ``AdalflowRAGPipeline``. It initializes the pipeline,
+adds documents from a directory, and generates responses for a list of user queries. The function is generic
+and can accommodate various LLM API clients, such as GroqAPIClient or OpenAIClient, highlighting the pipeline's
+flexibility and modularity.
+
+
+.. code-block:: python
+
+    def run_rag_pipeline(model_client, model_kwargs, documents, queries):
+
+        # Example usage of RAG pipeline
+        rag_pipeline = AdalflowRAGPipeline(
+            model_client=model_client,
+            model_kwargs=model_kwargs,
+            top_k_retrieval=1,  # Retrieve top 1 most relevant chunks
+            max_context_tokens=800  # Limit context to 1500 tokens
+        )
+
+        # Add documents from a directory of text files
+        rag_pipeline.add_documents_from_directory(documents)
+
+        # Generate responses
+        for query in queries:
+            print(f"\nQuery: {query}")
+            response = rag_pipeline.generate_response(query)
+            print(f"Response: {response}")
+
+
+This block provides an example of running the pipeline with different models and queries. It specifies:
+
+- The document directory containing the text files.
+- Example queries about topics such as the "Crystal Cavern" and "rare trees in Elmsworth."
+- Configuration for Groq and OpenAI model parameters, including the model type, temperature, and token limits.
+
+.. code-block:: python
+
+    documents = '../../tutorials/assets/documents'
+
+    queries = [
+        "What year was the Crystal Cavern discovered?",
+        "What is the name of the rare tree in Elmsworth?",
+        "What local legend claim that Lunaflits surrounds?"
+    ]
+
+    groq_model_kwargs = {
+        "model": "llama-3.2-1b-preview",  # Use 16k model for larger context
+        "temperature": 0.1,
+        "max_tokens": 800,
+    }
+
+    openai_model_kwargs = {
+        "model": "gpt-3.5-turbo",
+        "temperature": 0.1,
+        "max_tokens": 800,
+    }
+    # Below example shows that adalflow can be used in a genric manner for any api provider
+    # without worrying about prompt and parsing results
+    run_rag_pipeline(GroqAPIClient(), groq_model_kwargs, documents, queries)
+    run_rag_pipeline(OpenAIClient(), openai_model_kwargs, documents, queries)
+
+The example emphasizes that ``AdalflowRAGPipeline`` can interact seamlessly with multiple API providers,
+enabling integration with diverse LLMs without modifying the core logic for prompt construction or
+response parsing.
+
+
+.. admonition:: API reference
+   :class: highlight
+
+   - :class:`utils.setup_env`
+   - :class:`core.types.ModelType`
+   - :class:`components.model_client.OpenAIClient`
+   - :class:`components.model_client.GroqAPIClient`