LlamaIndex + Ertas

Use Ertas-trained models with LlamaIndex to build production-grade RAG applications that understand your domain-specific documents and data sources.

Overview

LlamaIndex (formerly GPT Index) is a data framework purpose-built for connecting large language models to external data sources. It provides robust abstractions for data ingestion, indexing, retrieval, and query engines that go far beyond simple vector search. LlamaIndex supports over 160 data connectors — from PDFs and databases to Slack, Notion, and Google Drive — making it straightforward to build RAG applications over enterprise knowledge bases.

What sets LlamaIndex apart is its focus on structured data retrieval and query planning. Rather than dumping raw chunks into a prompt, LlamaIndex can decompose complex queries into sub-queries, route them to different indexes, and synthesize coherent answers from multiple data sources. When paired with an Ertas-trained model that already understands your domain terminology, this structured retrieval approach dramatically reduces hallucinations and improves answer precision on domain-specific questions.

How Ertas Integrates

Ertas-trained models integrate with LlamaIndex at two critical points in the RAG pipeline: the query engine and the response synthesizer. After fine-tuning a model in Ertas Studio on your domain data, you deploy it to any OpenAI-compatible endpoint and configure LlamaIndex to use it as the primary LLM. Because the model already understands your industry jargon, abbreviations, and reasoning patterns, the retrieval-augmented responses are significantly more accurate than those from a generic model.

Ertas Hub provides pre-built LlamaIndex configuration templates for common use cases like legal document Q&A, medical literature review, and financial report analysis. These templates include optimized chunking strategies, embedding model recommendations, and prompt templates that align with the chat formats used during Ertas fine-tuning. This end-to-end alignment — from training data format to retrieval prompt structure — is what makes fine-tuned RAG pipelines outperform generic setups by wide margins.

Getting Started

1
Fine-tune a domain model in Ertas Studio
Train a model on your domain-specific Q&A pairs or document corpora using Ertas Studio. The fine-tuned model will serve as the reasoning backbone of your LlamaIndex pipeline.
2
Deploy the model to an inference endpoint
Export the model in GGUF format and serve it via Ollama, vLLM, or Ertas Cloud. LlamaIndex supports any OpenAI-compatible API as a backend.
3
Ingest and index your documents
Use LlamaIndex data connectors to load your documents, then build vector or keyword indexes using your preferred embedding model and vector store.
4
Configure the query engine with your Ertas model
Point LlamaIndex's query engine to your Ertas-trained model endpoint. Use the prompt templates from Ertas Hub to ensure training-inference prompt alignment.
5
Deploy and monitor the RAG application
Serve the LlamaIndex application via a REST API or chat interface. Use Ertas Cloud monitoring to track inference quality and identify areas for model improvement.

python

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
from llama_index.llms.openai_like import OpenAILike

# Connect to your Ertas-trained model via Ollama
llm = OpenAILike(
    api_base="http://localhost:11434/v1",
    model="ertas-finance-7b",
    api_key="not-needed",
    is_chat_model=True,
    temperature=0.1,
)

# Load and index your documents
documents = SimpleDirectoryReader("./financial_reports").load_data()
index = VectorStoreIndex.from_documents(documents)

# Query with your fine-tuned model
query_engine = index.as_query_engine(llm=llm)
response = query_engine.query(
    "What was the year-over-year revenue growth in Q3?"
)
print(response)

Build a RAG pipeline over financial documents using an Ertas-trained model with LlamaIndex.

Benefits

Domain-trained models produce more accurate RAG responses than generic LLMs
Support for 160+ data connectors covers virtually any enterprise data source
Structured query decomposition handles complex multi-part questions
All inference stays local or in your VPC for data privacy compliance
Pre-built templates from Ertas Hub accelerate LlamaIndex pipeline setup
Continuous improvement loop: feed RAG failures back into Ertas Studio for retraining