Custom-LLM-Chat / README.md
Bhaskar Ram
feat: apply all 15 upgrades — BGE embeddings, cosine FAISS, streaming LLM, tenacity retry, dotenv, Dockerfile, tests
a465955
---
title: Kerdos AI Custom LLM Chat (Demo)
emoji: 🤖
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: "6.6.0"
app_file: app.py
pinned: false
license: mit
tags:
- rag
- document-qa
- enterprise
- llama
- faiss
- gradio
- nlp
- question-answering
- kerdos
- private-llm
- demo
---
# 🤖 Kerdos AI — Custom LLM Chat & Document Q&A
> **A Demo Product by [Kerdos Infrasoft Private Limited](https://kerdos.in)**
> Upload your company documents. Ask questions. Get answers — strictly from your data.
> ⚠️ **This is a Demo Version.** Features, model selection, and customisation are limited.
> The full product will support **private, on-premise LLM deployments** tailored to your organisation.
---
## 🚀 About Kerdos Infrasoft
**Kerdos Infrasoft Private Limited** is an Indian technology startup incorporated in December 2023, headquartered in **Bengaluru, Karnataka, India** (CIN: U62099KA2023PTC182869). We specialise in:
- 🧠 Artificial Intelligence & Machine Learning
- 🔗 Blockchain Development
- 💻 Custom Software Development
- ☁️ Infrastructure & Digital Consulting
🌐 **Website**: [https://kerdos.in](https://kerdos.in)
📞 **Contact**: [https://kerdos.in/contact](https://kerdos.in/contact)
✉️ **Email**: [partnership@kerdos.in](mailto:partnership@kerdos.in)
---
## 💰 Seeking Investment & Partnerships
We are actively **seeking investment and strategic partnerships** to build the **fully customisable enterprise edition**, which will include:
- 🔒 **Private LLM Hosting** — on-premise or private-cloud deployments
- 🎛️ **Custom Model Fine-tuning** — trained on your domain data
- 🛡️ **Data Privacy Guarantees** — your data never leaves your environment
- 🏷️ **White-label Deployments** — fully branded for your organisation
**Interested?** Reach out at [partnership@kerdos.in](mailto:partnership@kerdos.in) or visit [kerdos.in/contact](https://kerdos.in/contact)
---
## ✨ Features (Demo)
| Feature | Details |
| ----------------------------- | ----------------------------------------------------------------- |
| 📄 **Multi-format ingestion** | PDF, DOCX, TXT, MD, CSV |
| 🧠 **Open-source LLM** | `meta-llama/Llama-3.1-8B-Instruct` via HF Inference API |
| 🔒 **Strictly grounded** | Answers only from your uploaded documents |
| 📦 **Multi-document** | Upload and query across multiple files simultaneously |
| 💬 **Multi-turn chat** | Maintains conversation context across questions |
| ⚡ **Fast** | CPU-friendly embeddings (`BAAI/bge-small-en-v1.5` + FAISS cosine) |
| 🔑 **Secure** | Files processed in-session only — never stored permanently |
---
## 🚀 How to Use
### On Hugging Face Spaces
1. Upload your documents (PDF, DOCX, TXT) using the left panel
2. Click **Index Documents**
3. Enter your [Hugging Face API token](https://huggingface.co/settings/tokens) _(Write access required for Llama 3)_
4. Ask questions in the chat!
### Self-Hosted / Local
```bash
git clone https://huggingface.co/spaces/kerdosdotio/Custom-LLM-Chat
cd Custom-LLM-Chat
pip install -r requirements.txt
HF_TOKEN=hf_your_token python app.py
```
---
## 🏗️ Architecture
```
User Uploads Files
Document Parser (PDF / DOCX / TXT)
Text Chunking (512 chars, 64 overlap)
Embeddings (BAAI/bge-small-en-v1.5)
FAISS Vector Index (cosine similarity, in-memory)
User Question → Similarity Search → Top-K Chunks
Llama 3.1 8B — answers ONLY from retrieved chunks
Response + Source Citations
```
---
## 🔧 Tech Stack
- **UI**: [Gradio](https://gradio.app)
- **LLM**: `meta-llama/Llama-3.1-8B-Instruct`
- **Embeddings**: `BAAI/bge-small-en-v1.5` (cosine similarity via FAISS)
- **Vector Store**: [FAISS](https://github.com/facebookresearch/faiss)
- **Document Parsing**: PyMuPDF, python-docx
---
## 💼 Use Cases
- **Customer Support**: Index your product manuals, FAQs, and policies
- **HR & Legal**: Query employee handbooks, contracts, and compliance docs
- **Sales Enablement**: Search product specs, case studies, and pricing docs
- **IT Helpdesk**: Query runbooks, troubleshooting guides, and SOPs
---
## 🔐 Privacy
- Uploaded documents are **processed in-memory** and **not stored** after your session ends
- For persistent storage or on-premise deployment with full privacy, [contact us](https://kerdos.in/contact)
---
## 📄 License
MIT License — free for commercial and private use.
---
_© 2024–2026 Kerdos Infrasoft Private Limited | Bengaluru, Karnataka, India_