This project presents a retrieval-augmented question-answering (QA) system designed to extract and summarize information from user-provided PDF documents. Leveraging LangChain and OpenAI's GPT-4, the system processes PDF text into semantic chunks, embeds them using OpenAI's models, and retrieves contextually relevant passages via a FAISS vector database. A custom prompt template guides the language model to generate concise, accurate answers from retrieved content. The pipeline ensures efficient handling of large documents while prioritizing contextual relevance and brevity in responses.
Text Extraction: PyPDF2 extracts raw text from PDFs, concatenating pages into a single string.
Text Chunking: A RecursiveCharacterTextSplitter divides text into manageable chunks (default: 1000 characters, 200-character overlap) to preserve semantic coherence.
Vector Database: OpenAI embeddings convert chunks into vectors, stored in FAISS for fast similarity searches. The retriever fetches the top 6 most relevant chunks per query.
Answer Generation: A LangChain pipeline combines the retriever, a custom QA prompt, and GPT-4. The prompt instructs the model to return concise answers (≤3 sentences) or admit uncertainty. The chain processes inputs dynamically, merging retrieved chunks into context for summarization.
The system successfully answers user questions by synthesizing retrieved PDF content. Key outcomes include:
Efficient chunking and embedding enable scalable processing of lengthy documents.
FAISS retrieves contextually relevant passages, improving answer accuracy.
GPT-4 generates succinct, coherent summaries, adhering to prompt constraints (e.g., brevity, uncertainty handling).
Testing confirms robustness across diverse PDF formats and question types, though performance hinges on input document quality.
This approach demonstrates the viability of retrieval-augmented generation (RAG) for domain-specific QA tasks without pre-training.
There are no models linked
There are no datasets linked
