What if your documents could speak back?
Synapse transforms static files into conversational companions.

Instead of scrolling endlessly, you ask questions like you would to a colleague:

• “Summarize section 2 in three lines.”
• “Compare results from table 3 with the conclusion.”
• “What assumptions are hidden in the methodology?”

Behind this, the system runs Retrieval-Augmented Generation (RAG):

1. Splits the document into small chunks 📄
2. Embeds them into a vector space 🌐 (where meaning is geometry)
3. Retrieves the closest knowledge for your query 🔍
4. Feeds both your question + retrieved chunks into an LLM 🧠

The answer is grounded, contextual, and transparent.
Not AI hallucination — but AI amplified by retrieval.

• 📄 PDF Uploads → Any report, research, policy, or notes.
• 💬 Natural Language Q&A → Just ask, no syntax needed.
• 🔍 Source Transparency → Exact page + snippet citations.
• ⚡ Local & Fast → Privacy-first, no external API calls.
• 🎨 Elegant UI → Next.js + Tailwind design, sleek & minimal.
• 🧠 Configurable AI → Swap models, tweak chunk size, tune temperature.

Here’s how Synapse works under the hood:

        ┌──────────────┐
        │   PDF Upload │ 📄
        └───────┬──────┘
                │
        ┌───────▼─────────┐
        │   Preprocessor  │ ✂️
        │ (chunk + clean) │
        └───────┬─────────┘
                │
        ┌───────▼─────────┐
        │  Embeddings     │ 🔮
        │ (Vectorization) │
        └───────┬─────────┘
                │
        ┌───────▼─────────┐
        │  Vector DB      │ 📊
        │   (Chroma)      │
        └───────┬─────────┘
                │
        ┌───────▼─────────┐
 Query  │   Retriever     │ 🔍
  💬 ──▶│ (find relevant) │
        └───────┬─────────┘
                │
        ┌───────▼─────────┐
        │   LLM Engine    │ 🧠
        │ (Phi-3 / LLaMA) │
        └───────┬─────────┘
                │
        ┌───────▼─────────┐
        │  Response +     │ ✨
        │  Citations      │
        └─────────────────┘

Synapse was deployed using ngrok, tunneling the local FastAPI backend to the web.
This meant the app was accessible publicly through a temporary URL — but only as long as my system was running.

⚡ It wasn’t production-grade hosting, but it made Synapse come alive for real-time demos.
A reminder that AI isn’t just about models — it’s about the entire pipeline, from embeddings to deployment.

For long-term usage, consider deploying with Docker + cloud hosting (AWS, GCP, Railway, etc.).

• Node.js ≥ 18
• Python ≥ 3.10
• Ollama installed
• 4GB+ RAM

git clone https://github.com/muskan-khushi/Synapse
cd synapse

cd server
python -m venv venv
source venv/bin/activate   # (Mac/Linux)
.\venv\Scripts\activate    # (Windows)

pip install -r requirements.txt

cd ../client
npm install

# Download Phi-3 Mini (3.9 GB) from HuggingFace
cp phi3.Modelfile.example phi3.Modelfile
# Update file path to gguf model
ollama create phi3-local -f phi3.Modelfile

ollama pull phi3:mini

Backend (Terminal 1):

cd server
uvicorn main:app --reload --port 8000

Frontend (Terminal 2):

cd client
npm run dev

🌐 Visit → http://localhost:3000

Ask your documents like a friend:

• “Summarize this in 5 bullet points.”
• “Which section mentions neural networks?”
• “Explain section 4 as if I’m 12 years old.”
• “Compare methodology with conclusion.”

Every answer → with citations & context 🔍.

Tune your AI in server/.env:

MODEL_NAME=phi3-local
OLLAMA_BASE_URL=http://localhost:11434
CHUNK_SIZE=1000
TEMPERATURE=0.1

• CHUNK_SIZE: how much text per vector slice
• TEMPERATURE: creativity of responses (0 = factual, 1 = creative)

Synapse is not just code — it’s a belief.
That documents are meant to be alive, not static.
That AI should be transparent, not black-boxed.
That embeddings + neural nets can feel almost poetic in how they connect meaning across dimensions.

It is research, but it is also art.
A bridge between mathematical rigor and dreamy interaction ✨.

MIT License — free to use, remix, and explore.