RAG-Based Technical Writing Assistant

This project is a scalable, style-guide-compliant rewriting assistant that uses Retrieval-Augmented Generation (RAG) to enforce writing rules like:

• Passive voice elimination
• Wordiness reduction
• Clarity and conciseness
• Minimalism
• Contractions and tone adjustments

It uses FAISS for vector-based retrieval, Hugging Face embeddings, and Gemma 7B via Ollama as the local LLM.

---

Directory Structure

.
├── rag_pipeline.py           # Main logic that retrieves rules + rewrites text
├── styleguide_ingest.py      # One-time script to embed .md style rules into FAISS DB
├── style_guide_rules/        # Folder containing individual .md files per rule
│   ├── passive_voice.md
│   ├── contractions.md
│   ├── minimalism.md
│   └── ...
├── vector_db/                # Auto-generated FAISS vector store
└── app.py                    # Gradio interface (optional)

---

How It Works

1. You write your style rules in .md format inside style_guide_rules/
2. You run styleguide_ingest.py to build/update your vector_db
3. When a user submits text, rag_pipeline.py:
   • Retrieves relevant rule chunks from the vector DB
   • Feeds them to the Gemma 7B model via Ollama
   • Gets a rewritten output that complies with those rules

---

Add New Style Rules

1. Add a new .md file in style_guide_rules/ (e.g., clarity.md)
2. Run:

python styleguide_ingest.py

No need to update any Python logic!

---

Requirements

• Python 3.10+
• ollama with gemma:7b model pulled locally
• langchain-community
• huggingface_hub, faiss-cpu, gradio, transformers

Install requirements:

pip install -r requirements.txt

---

Test It

After embedding style rules:

python app.py

Use the Gradio UI to paste passive voice–heavy or wordy text. You’ll get a clean, active, concise rewrite.

---

Example Rule: passive_voice.md

**Passive:** The file was reviewed by the QA team.  
**Active:** The QA team reviewed the file.

---

Model

You can change gemma:7b in rag_pipeline.py to:

• mistral:7b
• llama3:8b
• Or any other local model pulled into Ollama

---

Credits

Built with by Gaurav using:

• LangChain
• FAISS
• Hugging Face
• Microsoft Writing Style Guide