---

license: apache-2.0

KerdosAI — Universal LLM Training & RAG Agent

Enterprise-grade LLM Training + Retrieval-Augmented Generation (RAG) toolkit.
Fine-tune any HuggingFace model on your data, then deploy it with a full document Q&A chat interface — locally or on cloud.

---

What's New in v0.2.0

• 🆕 kerdosai.rag submodule — full RAG pipeline (document loading → FAISS indexing → LLM answering)
• 🆕 KnowledgeBase — high-level API to index PDF/DOCX/TXT/MD/CSV files
• 🆕 RAGAgent — streaming and blocking chat with conversation history
• 🆕 deployment_type="gradio-rag" — launch the HuggingFace Space UI locally with one line
• 🆕 kerdosai rag-chat CLI command — on-premise RAG UI in one command
• Updated dependencies: faiss-cpu, sentence-transformers, PyMuPDF, python-docx, gradio, huggingface-hub, tenacity

---

Installation

# Standard install
pip install kerdosai

# With all optional extras
pip install "kerdosai[all]"

Requirements: Python 3.8+ · PyTorch 2.0+ · CUDA-compatible GPU (recommended for training)

---

Quick Start

1. RAG — Document Q&A (no GPU needed)

from kerdosai.rag import KnowledgeBase, RAGAgent

# Index your documents (PDF, DOCX, TXT, MD, CSV)
kb = KnowledgeBase()
kb.index_documents(["handbook.pdf", "faq.docx", "policy.txt"])
print(f"Indexed {kb.chunk_count} chunks from {kb.indexed_sources}")

# Chat with your documents (HF Inference API — no GPU required)
agent = RAGAgent(hf_token="hf_...", knowledge_base=kb)

# Blocking answer
print(agent.chat("What is the leave policy?"))

# Streaming answer
for partial in agent.chat_stream("Summarise the refund section."):
    print(partial, end="\r")

2. Launch the Enterprise Chat UI Locally

from kerdosai.deployer import Deployer
from kerdosai.rag import KnowledgeBase

kb = KnowledgeBase()
kb.index_documents(["report.pdf"])

deployer = Deployer(model=None, tokenizer=None)
deployer.deploy(
    deployment_type="gradio-rag",
    host="0.0.0.0",
    port=7860,
    hf_token="hf_...",       # or set HF_TOKEN env var
    knowledge_base=kb,
)
# → Open http://localhost:7860

3. CLI — One-Command RAG Chat Server

# Pre-index files and open the Gradio UI
kerdosai rag-chat \
  --data handbook.pdf faq.docx policy.txt \
  --hf-token hf_... \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --port 7860

# Use HF_TOKEN env var instead of --hf-token
export HF_TOKEN=hf_...
kerdosai rag-chat --data ./company_docs/ --port 7860

4. Fine-Tune a Model (Training Pipeline)

from kerdosai import KerdosAgent

agent = KerdosAgent(
    base_model="meta-llama/Llama-3.1-8B",
    training_data="data/training.csv",
)

metrics = agent.train(epochs=3, batch_size=4, learning_rate=2e-5)
agent.save("./my-finetuned-model")
print(metrics)

5. Deploy Fine-Tuned Model as REST API

from kerdosai import KerdosAgent

agent = KerdosAgent.load("./my-finetuned-model")
agent.deploy(deployment_type="rest", host="0.0.0.0", port=8000)
# POST /generate → {"text": "...", "max_length": 200}

---

RAG Module Reference

KnowledgeBase

from kerdosai.rag import KnowledgeBase

kb = KnowledgeBase(
    embedding_model="BAAI/bge-small-en-v1.5",  # SentenceTransformer model
    chunk_size=512,                              # Max chars per chunk
    chunk_overlap=64,                            # Overlap between chunks
    min_score=0.30,                              # Min cosine similarity threshold
)

kb.index_documents(["doc1.pdf", "doc2.docx"])   # Add documents (duplicate-safe)
kb.index_documents(["doc3.txt"])                 # Incrementally add more

results = kb.search("What is the refund policy?", top_k=5)
# → [{"source": "doc1.pdf", "text": "...", "score": 0.87}, ...]

print(kb.chunk_count)       # Total indexed chunks
print(kb.indexed_sources)   # Set of indexed filenames
kb.clear()                  # Reset index

RAGAgent

from kerdosai.rag import RAGAgent

agent = RAGAgent(
    hf_token="hf_...",                              # HF API token
    model="meta-llama/Llama-3.1-8B-Instruct",       # LLM model ID
    top_k=5,                                         # Chunks retrieved per query
    embedding_model="BAAI/bge-small-en-v1.5",        # Embedding model
)

agent.index(["report.pdf", "handbook.docx"])         # Index documents

# Blocking
reply = agent.chat("What are the payment terms?")

# Streaming
for partial in agent.chat_stream("Summarise section 3."):
    print(partial, end="\r")

agent.reset_history()                                # Clear conversation

Low-Level RAG API

from kerdosai.rag import (
    load_documents,    # Parse files → [{"source", "text"}]
    build_index,       # List[dict] → VectorIndex
    add_to_index,      # Incrementally extend a VectorIndex
    retrieve,          # VectorIndex + query → top-K chunks
    answer_stream,     # Chunks + HF token → streaming generator
    answer,            # Chunks + HF token → full string
    VectorIndex,       # Dataclass: chunks + faiss index + embedder
)

docs     = load_documents(["report.pdf", "policy.docx"])
index    = build_index(docs, embedding_model="BAAI/bge-small-en-v1.5")
chunks   = retrieve("What is the refund policy?", index, top_k=5)
response = answer("What is the refund policy?", chunks, hf_token="hf_...")

Supported file types: .pdf (PyMuPDF) · .docx (python-docx, incl. tables) · .txt · .md · .csv

---

Training API Reference

KerdosAgent

Method	Description
KerdosAgent(base_model, training_data, device=None)	Initialize with a HuggingFace model ID and path to training data
.train(epochs, batch_size, learning_rate, **kwargs)	Fine-tune the model; returns metrics dict
.deploy(deployment_type, host, port, **kwargs)	Deploy as REST / Docker / Kubernetes / Gradio-RAG
.save(output_dir)	Save model + tokenizer to disk
.load(model_dir)	Class method — load a saved model

Trainer

Method	Description
Trainer(model, tokenizer, device, use_wandb=True)	Initialize trainer
.train(dataset, epochs, batch_size, learning_rate, ...)	Run HuggingFace training loop
.evaluate(dataset, batch_size)	Evaluate and return eval_loss

DataProcessor

Method	Description
DataProcessor(data_path, max_length=512, text_column="text")	Initialize
.prepare_dataset(tokenizer=None)	Load, clean, tokenize → HuggingFace Dataset
.validate_data()	Check data quality and print warnings

Supported training data formats: .csv (with text column) · .json (list of objects with text key)

---

CLI Reference

kerdosai <command> [options]

Commands:
  train        Fine-tune a model on custom data
  deploy       Deploy a trained model
  rag-chat     Launch a RAG document Q&A UI (no local GPU needed)

kerdosai train

kerdosai train \
  --model meta-llama/Llama-3.1-8B \
  --data ./data/training.csv \
  --output ./my-model \
  --epochs 3 \
  --batch-size 4 \
  --learning-rate 2e-5

kerdosai deploy

# REST API
kerdosai deploy --model-dir ./my-model --type rest --port 8000

# Gradio RAG UI (with a fine-tuned local model)
kerdosai deploy --model-dir ./my-model --type gradio-rag --hf-token hf_... --port 7860

# Docker container
kerdosai deploy --model-dir ./my-model --type docker

kerdosai rag-chat

kerdosai rag-chat \
  --data company.pdf handbook.docx faq.txt \
  --hf-token hf_... \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --embedding-model BAAI/bge-small-en-v1.5 \
  --host 0.0.0.0 \
  --port 7860

---

Deployment Options

Type	Command	Description
rest	deploy(deployment_type="rest")	FastAPI REST API on POST /generate
gradio-rag	deploy(deployment_type="gradio-rag")	Full Kerdos RAG Chat UI (Gradio)
docker	deploy(deployment_type="docker")	Build & run Docker container
kubernetes	deploy(deployment_type="kubernetes")	Generate K8s YAML manifests

---

Architecture

kerdosai/
├── __init__.py          # KerdosAgent, Trainer, Deployer, DataProcessor, KnowledgeBase, RAGAgent
├── agent.py             # KerdosAgent — orchestrates training + deployment
├── trainer.py           # Trainer — HuggingFace training loop + W&B logging
├── deployer.py          # Deployer — REST / Gradio-RAG / Docker / Kubernetes
├── data_processor.py    # DataProcessor — CSV/JSON loading, tokenization
├── cli.py               # CLI — train / deploy / rag-chat commands
└── rag/
    ├── __init__.py          # Public RAG API surface
    ├── document_loader.py   # PDF / DOCX / TXT / MD / CSV parser
    ├── embedder.py          # FAISS + SentenceTransformer index builder
    ├── retriever.py         # Top-K cosine-similarity retrieval
    ├── chain.py             # HF Inference API streaming + blocking answer
    ├── knowledge_base.py    # KnowledgeBase high-level class
    └── rag_agent.py         # RAGAgent — chat + history management

---

HuggingFace Space Demo

Try the live RAG demo at 👉 huggingface.co/spaces/kerdosdotio/Custom-LLM-Chat

Upload any PDF, DOCX, or TXT file and ask questions. The AI answers only from your documents — never from outside knowledge.

The full kerdosai package lets you run this same UI privately on your own servers with kerdosai rag-chat.

---

Environment Variables

Variable	Description	Default
HF_TOKEN	HuggingFace API token for inference	—
LLM_MODEL	LLM model ID for generation	meta-llama/Llama-3.1-8B-Instruct

---

Real-World Applications

Healthcare

• Clinical documentation automation
• Patient Q&A from medical records
• HIPAA-compliant private deployment

Financial Services

• Policy & compliance Q&A
• Risk report summarisation
• Private on-premise data security

Legal

• Contract review and clause extraction
• Case research from uploaded precedents
• Confidential document handling

Enterprise Internal Tools

• HR handbook chatbot
• IT knowledge base
• Onboarding document Q&A

---

Requirements

Python >= 3.8
torch >= 2.0.0
transformers >= 4.30.0
faiss-cpu >= 1.7.4
sentence-transformers >= 2.2.2
PyMuPDF >= 1.22.5
python-docx >= 0.8.11
gradio >= 4.0.0
huggingface-hub >= 0.28.0
tenacity >= 8.2.0

Full list in requirements.txt.

---

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

---

License

MIT License — see LICENSE for details.

---

About Kerdos Infrasoft

Kerdos Infrasoft Private Limited · CIN: U62099KA2023PTC182869
🌐 kerdos.in · 📬 partnership@kerdos.in

We are actively seeking investment & partnerships to build the fully customisable enterprise edition — including private LLM hosting, custom model fine-tuning, data privacy guarantees, and white-label deployments.

---

Citation

@software{kerdosai2024,
  title  = {KerdosAI: Universal LLM Training & RAG Agent},
  author = {Kerdos Infrasoft Private Limited},
  year   = {2024},
  url    = {https://github.com/bhaskarvilles/kerdosai},
  note   = {v0.2.1}
}