Pikachu – Sanchit's AI Assistant

A RAG + LLM personal chatbot API deployed on Apache Server. Built by Sanchit Minocha as an end-to-end NLP engineering project.

---

What is Pikachu?

Pikachu is Sanchit's personal AI assistant — it answers questions about Sanchit Minocha: his research, career, projects, publications, opinions, and personality. It is deployed as a REST API and integrated as a chatbot on his personal website.

Ask it things like:

• "What is Sanchit's most impactful project?"
• "Tell me about the GRILSS dataset."
• "What publications has Sanchit written in 2025?"
• "What are his hobbies?"

---

How It Works

Pikachu uses Retrieval-Augmented Generation (RAG). When you ask a question:

1. The query is cleaned — person's name and question scaffolding are stripped so the embedding focuses on the actual topic
2. Follow-up detection: if the query contains a dangling pronoun ("it", "that", "those") or is a continuation phrase ("yes", "tell me more", "other than that?"), the last user message and a snippet of the last assistant response are prepended — this grounds the embedding in prior context while steering retrieval toward new content not yet covered
3. The augmented query is embedded with BAAI/bge-base-en-v1.5 (ONNX, 768-dim) and ChromaDB finds the most relevant chunks via cosine similarity
4. Chunks scoring below the similarity threshold are discarded — the LLM receives an explicit "no information found" marker rather than empty context
5. Surviving chunks are injected as numbered, labelled context into the LLM prompt
6. The LLM generates a response grounded strictly in that context; the fallback "I don't know" phrase only fires when the context block is truly empty

LLM backends (with automatic fallback in order):

1. Groq — llama-3.3-70b-versatile (fast, free tier)
2. Cerebras — llama3.3-70b (fast, free — CEREBRAS_API_KEY)
3. OpenRouter — openai/gpt-oss-20b:free (free models — OPENROUTER_API_KEY)
4. HuggingFace — meta-llama/Llama-3.2-3B-Instruct (requires fine-grained token)
5. Ollama — local, no internet needed

For a deeper technical explanation → docs/architecture.md

---

API

POST /api/chat

curl -X POST http://localhost:5050/api/chat \
  -H "Content-Type: application/json" \
  -d '{
    "message": "Who is Sanchit Minocha?",
    "conversation_id": "abc-123",
    "top_k": 7,
    "max_tokens": 500,
    "model": "llama-3.3-70b-versatile"
  }'

Request fields:

Field	Type	Default	Description
message	string	required	The question to ask
conversation_id	string	auto-generated	Pass back the returned ID for multi-turn conversation
top_k	int	7	Number of RAG chunks to retrieve (max 10)
max_tokens	int	500	Maximum tokens in the response
model	string	from .env	Override the LLM (see accepted values below)

Accepted model values:

Value	Backend	Notes
llama-3.3-70b-versatile	Groq	Best quality — recommended default
llama-3.1-70b-versatile	Groq	Good quality
llama-3.1-8b-instant	Groq	Fast, lower accuracy
llama3-8b-8192	Groq	Groq-hosted Llama 3 8B
llama3-70b-8192	Groq	Groq-hosted Llama 3 70B
mixtral-8x7b-32768	Groq	Mixtral MoE
gemma2-9b-it	Groq	Google Gemma 2 9B
gemma-7b-it	Groq	Google Gemma 7B
llama3.3-70b	Cerebras	Free — set CEREBRAS_API_KEY in .env
llama3.1-70b	Cerebras	Free — fast 70B model
llama3.1-8b	Cerebras	Free — fast, smaller model
gpt-oss-120b	Cerebras	Free — large OSS model
qwen-3-235b-a22b-instruct-2507	Cerebras	Free — Qwen 3 235B MoE
openai/gpt-oss-20b:free	OpenRouter	Free — default; set OPENROUTER_API_KEY in .env
meta-llama/llama-3.2-3b-instruct:free	OpenRouter	Free — lightweight Llama
google/gemma-3-1b-it:free	OpenRouter	Free — lightweight Gemma
phi3:mini	Ollama (local)	Requires ollama pull phi3:mini
llama3.2:3b	Ollama (local)	Requires ollama pull llama3.2:3b
meta-llama/Llama-3.2-3B-Instruct	HuggingFace	Requires fine-grained token with inference permission

Response:

{
  "response": "Sanchit Minocha is a PhD researcher at the University of Washington...",
  "model": "llama-3.3-70b-versatile",
  "sources": ["profile.md", "personal_data.json"],
  "conversation_id": "abc-123",
  "timestamp": "2024-01-01T00:00:00Z"
}

Other Endpoints

Endpoint	Method	Description
GET /api/health	GET	LLM status, index size, backend availability
GET /api/info	GET	Assistant metadata and creator info
POST /api/rebuild-index	POST	Rebuild RAG index after editing knowledge files

---

Quick Start

git clone https://github.com/SanchitMinocha/Pikachu-AI-Assistant.git
cd sanchitai

python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

cp .env.example .env        # fill in your GROQ_API_KEY
python scripts/build_index.py
python app.py               # dev server on :5050

For full Apache deployment → docs/deployment.md

---

Testing

Browser (easiest)

Open test.html in any browser — it provides a full chat UI with model selector, source display, and health status. Works against the dev server or Apache:

# Dev server
open test.html        # macOS
xdg-open test.html    # Linux

# Or just drag-and-drop into your browser

The default base URL in the tester is http://localhost:5050. You can change it in the Settings panel at the top of the page.

Terminal — quick checks

The server uses a self-signed SSL certificate, so pass -k to skip certificate verification:

# Health check
curl -sk https://f-hossain-3.ce.washington.edu/scb/api/health | python3 -m json.tool

# Simple chat
curl -sk -X POST https://f-hossain-3.ce.washington.edu/scb/api/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "Tell me about RAT 3.0"}' | python3 -m json.tool

# Multi-turn conversation (reuse conversation_id)
curl -sk -X POST https://f-hossain-3.ce.washington.edu/scb/api/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "What about GRILSS?", "conversation_id": "test-1"}' \
  | python3 -m json.tool

# Try a specific model
curl -sk -X POST https://f-hossain-3.ce.washington.edu/scb/api/chat \
  -H "Content-Type: application/json" \
  -d '{"message": "What publications did Sanchit write in 2025?", "model": "llama-3.3-70b-versatile"}' \
  | python3 -m json.tool

# Rebuild the index after editing knowledge files
curl -sk -X POST https://f-hossain-3.ce.washington.edu/scb/api/rebuild-index | python3 -m json.tool

# Assistant info
curl -sk https://f-hossain-3.ce.washington.edu/scb/api/info | python3 -m json.tool

For local dev server (python app.py), drop the -k and use http://localhost:5050.

---

Knowledge Base

All knowledge about Sanchit lives in plain editable files:

data/
├── personal_data.json          ← hobbies, stories, opinions, FAQs — edit freely
├── knowledge_base/
│   ├── profile.md              ← career, education, skills
│   ├── projects.md             ← project details (split by ## section)
│   ├── about_ai.md             ← what this assistant is
│   ├── github.md               ← auto-fetched GitHub repos
│   └── Sanchit_CV.pdf          ← CV (auto-indexed as PDF)
└── website_pages/
    ├── publications.json       ← research publications (auto-indexed)
    └── portfolios.json         ← portfolio projects (auto-indexed)

Each ##-headed section in Markdown files becomes one chunk. Publications and portfolio entries each become their own self-contained chunks — so asking about a specific paper retrieves exactly that paper's context.

After editing any file, rebuild the index:

python scripts/build_index.py

For adding more data, managing publications, or adapting for a different person → docs/personalization.md

---

Security — What NOT to Commit

File	Why	What to commit instead
.env	Contains API keys (Groq, HuggingFace)	.env.example (placeholder keys)
data/processed/	Generated ChromaDB vector store	Rebuilt by build_index.py
data/models/	Downloaded ONNX model files (~90MB)	Auto-downloaded on first run
models/	Fine-tuned model weights (GBs)	—
logs/	Server logs	—

All of the above are already in .gitignore.

---

Project Structure

sanchitai/
├── app.py                    API entry point (Flask)
├── config.py                 All settings
├── wsgi.py                   Apache mod_wsgi entry point
├── requirements.txt
├── .env.example              Safe config template (commit this)
├── test.html                 Browser-based chat tester
│
├── data/
│   ├── personal_data.json    ← EDIT THIS to personalize
│   ├── knowledge_base/       Markdown + PDF knowledge files
│   └── website_pages/        ← publications.json + portfolios.json
│
├── src/
│   ├── rag/                  Embeddings + ChromaDB + retriever
│   ├── llm/                  Groq / Ollama / HuggingFace backends
│   └── data/                 Document loading and chunking
│
├── scripts/
│   ├── build_index.py        Build RAG vector index
│   ├── collect_web_data.py   Refresh GitHub data
│   └── fine_tune.py          Optional LoRA fine-tuning
│
├── docs/
│   ├── architecture.md       How it works (RAG pipeline explained)
│   ├── deployment.md         Apache deployment step-by-step
│   └── personalization.md    How to update data and retrain
│
└── apache/
    └── sanchit-ai.conf       Apache VirtualHost config

---

About the Creator

Sanchit Minocha is a PhD researcher at the University of Washington (Geospatial AI & Water Resource Management), IIT Roorkee Gold Medalist, and data scientist with 7+ years of experience building tools that use satellite data to manage global water resources.

• Website: sanchitminocha.github.io
• GitHub: github.com/SanchitMinocha
• LinkedIn: linkedin.com/in/sanchitminochaiitr

---

MIT License — fork and adapt for your own personal AI assistant.