A small Flask web app that turns a stack of PDFs into a grounded question-answering chatbot, with an optional voice loop. Upload a few documents, ask a question in the browser (or speak it), and the assistant retrieves the most relevant passage from your PDFs and asks GPT-3.5 to answer using only that passage.

Built September 2023 as a focused exercise in retrieval-augmented generation before "RAG" was the default acronym in every tutorial. Kept here as a reference implementation of the core loop without the framework weight of LangChain or LlamaIndex.

I wanted to see how far a deliberately minimal RAG stack could go for a single domain:

• Plain PyPDF2 for ingest.
• scikit-learn TfidfVectorizer plus cosine similarity for retrieval (no vector DB).
• A tightly-scoped system prompt that refuses to answer from outside the supplied passage.
• A Flask front-end with a chat widget and a "human-like typing" effect on the response.

An optional path uses sentence-transformers plus FAISS for dense retrieval, with a similarity check that verifies the model's answer is actually grounded in the source text before returning it (see retrieval_from_embeddings.py).

The voice loop (SpeechRecognition + pyttsx3) is included but commented out in the main flow, because the web UI ended up being the more useful surface.

PDF(s) ──> PDFProcessor ──> sentence list
                                 │
                                 ▼
                     TfidfVectorizer.fit_transform
                                 │
                                 ▼
   user question ──> cosine_similarity ──> best-match passage
                                                 │
                                                 ▼
                       OpenAI Chat Completions (gpt-3.5-turbo)
                       system prompt: "use only this passage"
                                                 │
                                                 ▼
                              answer ──> Flask UI / pyttsx3 TTS

Files:

• app.py: Flask routes (/, /chat, /upload).
• assistant.py: Orchestrator that wires PDF ingest, retrieval, and chat.
• pdf_processor.py: Extracts and tokenises PDF text into sentences.
• chatbot.py: TF-IDF retrieval plus OpenAI Chat Completions call.
• audio_processor.py: speech_recognition (Google Web Speech API) for STT, pyttsx3 (SAPI5 voice) for TTS.
• retrieval_from_embeddings.py: Alternate path using sentence-transformers embeddings, FAISS index, and a grounding-verification step.
• templates/index.html, static/: Flask UI plus a small chat widget.

Requires Python 3.10+ and an OpenAI API key. The voice features additionally require system audio libraries (see below).

git clone https://github.com/dparikh79/AI-Assistant.git
cd AI-Assistant

python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate

pip install flask PyPDF2 scikit-learn requests \
            SpeechRecognition pyttsx3 gTTS \
            sentence-transformers faiss-cpu

Create secrets.json in the project root (it is gitignored):

{
  "API_ENDPOINT": "https://api.openai.com/v1/chat/completions",
  "API_KEY": "sk-your-openai-key"
}

Edit config.py to point PATH_TO_BOOK at your own PDFs, or leave the list empty and upload PDFs through the web UI at runtime.

Run the web app:

python app.py
# open http://127.0.0.1:5000

Or run the terminal-only loop (no Flask):

python assistant.py

• macOS: brew install portaudio then pip install pyaudio.
• Debian/Ubuntu: sudo apt-get install portaudio19-dev espeak ffmpeg then pip install pyaudio.
• Windows: pyttsx3 uses SAPI5 out of the box; pip install pyaudio usually works on recent Python builds.

The voice calls are commented out in assistant.py::main. Uncomment the AudioProcessor.text_to_speech(...) and AudioProcessor.voice_to_text() lines to switch from typed input to spoken input.

This is a 2023-era sketch, not a product. Worth being upfront about what it is and is not:

• TF-IDF retrieval over sentence splits is fast but brittle on long, multi-page reasoning. The FAISS path in retrieval_from_embeddings.py is better but not wired into the Flask app.
• pyttsx3 is initialised with sapi5, so out-of-the-box TTS is Windows-only. On macOS or Linux, swap the driver (nsss or espeak) in audio_processor.py.
• The Google Web Speech backend used by SpeechRecognition is undocumented and unsuitable for production. A modern rewrite would use Whisper (local) or the OpenAI audio API.
• No wake word, no OS control (no app launching, no web search). It is a grounded Q&A loop, not a Siri replacement.
• Bare requests.post to OpenAI with no retry/backoff and no streaming; a try/except swallows errors and prints to stdout.
• The app.secret_key in app.py is a placeholder. Set it from an environment variable before exposing the app to anyone other than yourself.

If I rebuilt this in 2026 I would: swap PyPDF2 for pypdf or unstructured; replace TF-IDF with a small embedding model and an actual vector store; move the OpenAI call to the official SDK with streaming and structured output; use Whisper for STT; and put the grounding check from retrieval_from_embeddings.py in front of every answer, not just the alternate path.

MIT. See LICENSE.