📘 PDF Q&A Bot — Chunking & Embedding Playground
A Streamlit-powered application that allows you to upload a PDF, chunk its content, generate embeddings using multiple providers (Gemini, OpenAI, Cohere, HuggingFace), store them in a FAISS vector index, and ask natural language questions whose answers are retrieved from the PDF content.
📂 Upload any PDF and extract its text.
🪓 Multiple chunking methods:
- Simple split
- Character-based split
- Recursive character split
- Token-based split
🧠 Multiple embedding backends:
- Google Gemini
- OpenAI (yet to improve)
- HuggingFace Sentence Transformers (yet to improve)
- Cohere
🔍 Semantic search using FAISS for fast retrieval.
- Text Extraction → PDF text is extracted using
pypdf. - Chunking → Text is split into smaller segments to fit embedding model limits.
- Embedding → Selected embedding model converts chunks into numerical vectors.
- Indexing → FAISS stores the vectors for efficient similarity search.
- Retrieval → On a query, similar chunks are retrieved.
- Answer Generation → Model uses retrieved chunks as context to answer.
| Chunking Method | Description | Pros | Cons |
|---|---|---|---|
| Simple | Splits text into fixed-size chunks. | Fast, easy to implement. | May break sentences or lose context. |
| Character | Splits by a set number of characters. | Good for uniform text length. | Can split mid-word or mid-sentence. |
| Recursive | Splits respecting paragraphs/sentences first. | Preserves context, better for QA. | Slightly slower. |
| Token | Splits based on LLM token count. | Prevents token overflow errors. | Requires tokenizer, more complex. |
| Embedding Provider | Model Used | Strengths | Limitations |
|---|---|---|---|
| Google Gemini | models/embedding-001 |
High quality, multilingual, good semantic understanding. | Requires Google API key. |
| OpenAI | text-embedding-3-small |
High accuracy, widely supported in LangChain. | Paid API after free tier. |
| Hugging Face | sentence-transformers/all-MiniLM-L6-v2 |
Free, runs locally, no API needed. | Slightly slower on large docs. |
| Cohere | embed-english-light-v3.0 |
Fast, optimized for English. | Limited multilingual support. |
-
Chunk Size → The maximum number of characters or tokens in each chunk.
- Example:
chunk_size = 500→ each chunk is ~500 characters/tokens. - Effect: Larger chunks = more context per query, but risk hitting model token limits.
- Example:
-
Chunk Overlap → The number of characters/tokens that overlap between chunks.
- Example:
chunk_overlap = 50→ 50 characters/tokens repeated between consecutive chunks. - Effect: Higher overlap = better context retention between chunks, but more processing cost.
- Example: