====== RAGFlow ======

**RAGFlow** is an open-source RAG (Retrieval-Augmented Generation) engine developed by Infiniflow that specializes in deep document understanding through advanced parsing capabilities including OCR, table structure recognition, and document layout analysis. With over **76,000 GitHub stars**, it excels at handling complex documents that other RAG systems struggle with.

| **Repository** | [[https://github.com/infiniflow/ragflow|github.com/infiniflow/ragflow]] |
| **License** | Apache 2.0 |
| **Language** | Python |
| **Stars** | 76K+ |
| **Category** | RAG Engine |

===== Key Features =====

  * **Deep Document Understanding** -- Advanced parsing of complex PDFs with OCR, table extraction, and layout recognition
  * **DeepDoc Engine** -- Proprietary document analysis engine handling layout analysis, figure extraction, and rotation correction for scanned PDFs
  * **Table Structure Recognition (TSR)** -- YOLOv8-based fine-tuned models that outperform AWS Textract on complex tables
  * **Document Layout Recognition (DLR)** -- Identifies titles, paragraphs, figures, and multi-column layouts
  * **Visual Model Flexibility** -- Autonomous selection of visual models per task (OCR/TSR/DLR) to balance speed and accuracy
  * **Table of Contents Extraction** -- Uses LLMs during indexing to enable long-context RAG with structural context
  * **Multi-Format Support** -- PDF, DOCX, XLSX, CSV, images, emails, and plain text

===== Architecture =====

RAGFlow decouples data extraction from chunking (since v0.17.0), allowing independent selection of visual models for each processing task. The pipeline flows through ingestion, parsing, embedding, retrieval, and generation stages.

  * **Parser Layer** -- DeepDoc for advanced PDF handling, PlainParser for text, VisionParser for image-heavy documents
  * **Extraction Layer** -- Separate OCR, TSR, and DLR models that can be independently configured
  * **Chunking Layer** -- Structure-preserving strategies that maintain table rows with headers as self-contained units
  * **Retrieval Layer** -- Hybrid search combining vector similarity with structural context from table of contents
  * **Generation Layer** -- LLM-based answer generation with source attribution

<mermaid>
graph TB
    subgraph Input["Document Input"]
        PDF[PDF Documents]
        DOCX[Word / Excel]
        IMG[Images]
        TXT[Text / Email]
    end
    subgraph Parsing["Deep Document Parsing"]
        DeepDoc[DeepDoc Engine]
        OCR[OCR Module]
        TSR[Table Structure Recognition]
        DLR[Layout Recognition]
    end
    subgraph Processing["Processing Pipeline"]
        Chunk[Chunking Engine]
        TOC[TOC Extraction]
        Embed[Embedding Generator]
    end
    subgraph Storage["Storage Layer"]
        VDB[(Vector Database)]
        Meta[(Metadata Store)]
    end
    subgraph Query["Query Pipeline"]
        Retrieve[Hybrid Retrieval]
        Rerank[Reranking]
        Generate[LLM Generation]
    end
    Input --> Parsing
    DeepDoc --> OCR
    DeepDoc --> TSR
    DeepDoc --> DLR
    Parsing --> Processing
    Processing --> Storage
    Storage --> Query
    TOC --> Retrieve
</mermaid>

===== Document Parsing Details =====

RAGFlow's parsing capabilities are the core differentiator:

  * **PDF Parsing** -- The most robust capability; DeepDoc handles layout analysis, figure extraction, and auto-detects rotation (90/180/270 degrees) via OCR
  * **Table Extraction** -- Identifies complex tables including single-column and borderless layouts, outputting HTML/Markdown to preserve structure and relationships
  * **Spreadsheets** -- RAGFlowExcelParser with openpyxl/pandas extracts cell values while retaining structure as HTML
  * **DOCX** -- Uses python-docx for text and table extraction
  * **Images** -- OCR and optional Vision Language Model processing

===== Code Example =====

<code python>
import requests

RAGFLOW_API = "http://localhost:9380/api/v1"
API_KEY = "ragflow-your-api-key"
HEADERS = {"Authorization": f"Bearer {API_KEY}", "Content-Type": "application/json"}

# Create a knowledge base (dataset)
dataset = requests.post(f"{RAGFLOW_API}/datasets",
    headers=HEADERS,
    json={"name": "technical_docs", "chunk_method": "naive"}
).json()

dataset_id = dataset["data"]["id"]

# Upload a document
with open("complex_report.pdf", "rb") as f:
    upload = requests.post(
        f"{RAGFLOW_API}/datasets/{dataset_id}/documents",
        headers={"Authorization": f"Bearer {API_KEY}"},
        files={"file": f}
    ).json()

# Query the knowledge base with RAG
answer = requests.post(f"{RAGFLOW_API}/chats",
    headers=HEADERS,
    json={"question": "What were the Q3 revenue figures?",
          "dataset_ids": [dataset_id]}
).json()
print(answer["data"]["answer"])
</code>

===== References =====

  * [[https://github.com/infiniflow/ragflow|RAGFlow GitHub Repository]]
  * [[https://ragflow.io|RAGFlow Official Website]]
  * [[https://ragflow.io/docs/select_pdf_parser|RAGFlow PDF Parser Documentation]]
  * [[https://github.com/infiniflow/ragflow/blob/main/deepdoc/README.md|DeepDoc README]]

===== See Also =====

  * [[dify|Dify]] -- Agentic workflow platform with RAG capabilities
  * [[lightrag|LightRAG]] -- Knowledge graph-enhanced RAG
  * [[milvus|Milvus]] -- Vector database for RAG storage
  * [[chromadb|ChromaDB]] -- AI-native embedding database