LlamaIndex

LlamaIndex (formerly GPT Index) connects LLMs to your data. 2025 introduces Workflows, an event-driven way to build complex RAG pipelines.

When to Use

• RAG (Retrieval Augmented Generation): Indexing PDFs, Docs, SQL to chat with them.
• Structured Data: Querying SQL/Pandas with natural language (NLSQL).
• Agents: Building research agents that browse the web and summarize.

Core Concepts

Workflows

Event-driven architecture for agents. Replace DAGs with event listeners (@step).

Query Engine

High-level API (index.as_query_engine()) to ask questions.

Data Loaders (LlamaHub)

Connectors for Notion, Slack, Discord, PDF, etc.

Best Practices (2025)

Do:

• Use Workflows: They are harder to learn but easier to debug than monolithic engines.
• Use Hybrid Search: BM25 (Keyword) + Vector Search for best retrieval accuracy.
• Use Rerankers: Always rerank retrieved nodes (Cohere/BGE) before sending to LLM.

Don’t:

• Don’t dump raw text: Use “Node Parsers” to chunk data intelligently (Markdown, Semantic).

References

• LlamaIndex Documentation