Skip to content

Latest commit

 

History

History

haystack-tutorial

README.md

layout title nav_order has_children
default
Haystack Tutorial
23
true

Haystack: Deep Dive Tutorial

Project: Haystack — An open-source framework for building production-ready LLM applications, RAG pipelines, and intelligent search systems.

Stars License: Apache 2.0 Python

What Is Haystack?

Haystack is an open-source LLM framework by deepset for building composable AI pipelines. It provides a modular, component-based architecture that combines retrieval, generation, and evaluation into production-ready workflows. Haystack supports dozens of LLM providers, vector databases, and retrieval strategies out of the box.

Feature Description
Pipeline System Directed graph of components with typed inputs/outputs and automatic validation
RAG First-class retrieval-augmented generation with hybrid search (BM25 + embedding)
Multi-Provider OpenAI, Anthropic, Cohere, Google, Hugging Face, Ollama, and more
Document Stores In-memory, Elasticsearch, OpenSearch, Pinecone, Qdrant, Weaviate, Chroma, pgvector
Evaluation Built-in metrics (MRR, MAP, NDCG) and LLM-based evaluation components
Custom Components @component decorator for building reusable pipeline nodes with typed I/O

Current Snapshot (auto-updated)

Architecture Overview

graph TB
    subgraph Ingestion["Ingestion Pipeline"]
        FILES[File Converters]
        SPLIT[Document Splitter]
        EMBED_D[Document Embedder]
        WRITER[Document Writer]
    end

    subgraph Store["Document Stores"]
        MEM[In-Memory]
        ES[Elasticsearch]
        PG[pgvector]
        VEC[Pinecone / Qdrant / Weaviate]
    end

    subgraph Query["Query Pipeline"]
        EMBED_Q[Query Embedder]
        BM25[BM25 Retriever]
        EMB_RET[Embedding Retriever]
        JOINER[Document Joiner]
        RANKER[Ranker]
        PROMPT[Prompt Builder]
        GEN[Generator / LLM]
    end

    FILES --> SPLIT --> EMBED_D --> WRITER
    WRITER --> Store

    Store --> BM25
    Store --> EMB_RET
    EMBED_Q --> EMB_RET
    BM25 --> JOINER
    EMB_RET --> JOINER
    JOINER --> RANKER --> PROMPT --> GEN
Loading

Tutorial Structure

Chapter Topic What You'll Learn
1. Getting Started Setup Installation, first RAG pipeline, architecture overview
2. Document Stores Storage Store backends, indexing, preprocessing, multi-store patterns
3. Retrievers & Search Retrieval BM25, embedding, hybrid search, filtering, re-ranking
4. Generators & LLMs Generation Multi-provider LLMs, prompt engineering, streaming, chat
5. Pipelines & Workflows Composition Pipeline graph, branching, loops, serialization, async
6. Evaluation & Optimization Quality Retrieval metrics, LLM evaluation, A/B testing, optimization
7. Custom Components Extensibility @component decorator, typed I/O, testing, packaging
8. Production Deployment Operations REST API, Docker, Kubernetes, monitoring, scaling

Tech Stack

Component Technology
Language Python 3.9+
Pipeline Engine Custom directed graph with topological execution
Serialization YAML / JSON pipeline definitions
Embeddings Sentence Transformers, OpenAI, Cohere, Fastembed
Vector Search FAISS, Pinecone, Qdrant, Weaviate, Chroma, pgvector
Text Search Elasticsearch, OpenSearch, BM25 (in-memory)
LLM Providers OpenAI, Anthropic, Google, Cohere, Hugging Face, Ollama
API Layer Hayhooks (FastAPI-based pipeline serving)

Ready to begin? Start with Chapter 1: Getting Started.

Built with insights from the Haystack repository and community documentation.

Navigation & Backlinks

Full Chapter Map

  1. Chapter 1: Getting Started with Haystack
  2. Chapter 2: Document Stores
  3. Chapter 3: Retrievers & Search
  4. Chapter 4: Generators & LLMs
  5. Chapter 5: Pipelines & Workflows
  6. Chapter 6: Evaluation & Optimization
  7. Chapter 7: Custom Components
  8. Chapter 8: Production Deployment

Source References

Generated by AI Codebase Knowledge Builder