NoteLense

In Notelense, when a user uploads a PDF, the file is first parsed and its text is extracted. The extracted text is then chunked into smaller, meaningful sections so context can be preserved. Each chunk is passed through an embedding model Gemini embeddings to convert the text into numerical vectors. These vectors, along with metadata (such as page number, document ID, and session UUID), are stored in Pinecone, which acts as our vector database. During a chat, when the user asks a question, I embed the query, perform a semantic similarity search in Pinecone to fetch the most relevant chunks, and then feed those chunks along with the query into the LLM to generate a contextual, natural language response. This flow makes Notelense work like a personal research assistant, retrieving accurate information from PDFs on demand.