🤖 ShellSage

✨ Context-aware, low-latency command-line autocomplete powered by Trie, FAISS, and LLMs.

ShellSage is my attempt at an intelligent autocomplete engine for your terminal that combines the speed of Trie-based search, the semantic understanding of FAISS, and the reasoning capabilities of LLMs to provide blazing-fast, contextually relevant command suggestions.

NOTE: This is only supported in fish at the moment

---

🚀 Features

• ⚡ Low-latency Suggestions using Trie prefix search
• 🧠 Context-aware Search using FAISS semantic embeddings
• 🤖 LLM-Powered Completions to refine and personalize suggestions
• 📚 User Command History Integration
• 📃️ Hybrid Ranking Engine combining all sources
• 🧵 gRPC-based Streaming API for interactive clients
• 💾 FAISS Cache to reuse LLM results and reduce API latency
• 🔐 Modular, extensible engine built in Python

---

🏗️ Architecture

                 ┌─────────────┐
                 │   Terminal  │
                 └─────────────┘
                        │
               gRPC backend client
                        │
              ┌───────────────────┐
              │   ShellSage Core  │
              └───────────────────┘
   ┌────────────────────│──────────────────────┐
   │             │              │              │
TrieSearch   FaissSearch   LLMCompletion   FaissCache
(prefix)     (semantic)     (GPT-like)     (embedding cache)

---

⚙️ Installation

# Clone the repo
git clone https://github.com/your-username/shellsage.git
cd shellsage

# Set up virtual environment
python3 -m venv .venv
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Install ollama
<install ollama>

# Pull mistral
ollama pull mistral

---

🥪 Running the Server

python -m src.api.server

This will start the gRPC server on localhost:50051. It uses a thread pool executor and gracefully handles streamed prompts from clients.

---

🧠 Core Components

Component	Description
TrieSearch	Fast prefix matching from shell history
FaissSearch	Vector search using sentence embeddings
LLMCompletion	Calls local AI model to refine suggestions
FaissCache	Embedding-keyed cache to avoid redundant LLM calls
Ranker	Combines and scores all results
gRPC Server	Streams back autocomplete results to clients

---

📁 Project Structure

src/
├── api/                 # gRPC server & protobufs
├── prediction_engine/   # Core engine components
│   ├── core/            # Trie, FAISS, LLM logic
│   ├── data/            # History loader
│   └── ranking/         # Result ranking strategy
├── utils/               # Logging, helpers
├── protobuf/            # Generated gRPC code

---

📝 TODO

• Shell plugin for Bash/Zsh/Fish autocompletion
• Persistent command history across sessions
• Create a proper RAG for LLM (might increase latency)
• Session-aware completions

---

📄 License

MIT License © 2025 Harsh S.