RAGflow

Retrieval-Augmented Generation service for enterprise document indexing, semantic search, and knowledge graphs

ragflow

ragflowragvector-searchpgvectorembeddingsdocument-indexingknowledge-graph

Updated 2026-11-03

Verified 2026-11-03

The RAGflow service (cosmos-ragflow-v3) provides Retrieval-Augmented Generation (RAG) capabilities for enterprise document indexing and semantic search. It processes documents, generates embeddings, and stores them in PostgreSQL with pgvector for vector similarity search. It also builds knowledge graphs over indexed files and runs retrieval evaluations.

Architecture

Framework: FastAPI + SQLAlchemy; the webserver applies Alembic migrations (alembic upgrade head) when it starts
Vector Database: PostgreSQL + pgvector (VECTOR_STORE_TYPE default postgres)
Graph Store: Neo4j by default (GRAPH_STORE_TYPE)
Async Processing: Celery + Redis, with separate queues for bulk indexing (ragflow-indexing), interactive work (ragflow-interactive), and deep search (ragflow-search)
Embeddings: EMBEDDING_PROVIDER (default azure) with EMBEDDING_MODEL (default text-embedding-3-small)
Document Parsing: PDF (Docling, run in a subprocess), DOCX, PPTX, XLSX, and images including GIF (Tesseract OCR)
Dependencies: uv (pyproject.toml + uv.lock); the image runs uv sync into /app/.venv
Telemetry: Cloud deployments start each process under opentelemetry-instrument, which reports to Azure Monitor through APPLICATIONINSIGHTS_CONNECTION_STRING

Components

Component	Container	Description
API Server	`ragflow-web`	REST API for parsing, indexing, search, graphs, and evaluation
Worker	`ragflow-worker`	Celery worker for bulk document indexing
Interactive Worker	`ragflow-interactive-worker`	Celery worker for short interactive tasks
Search Worker	`ragflow-search-worker`	Celery worker for deep-search jobs

Processing Pipeline

RAGflow indexing and retrieval pipeline

Source

Loading diagram...

API Overview

Files

Method	Endpoint	Description
POST	`/api/v1/parse`	Parse a Storage file into text
POST	`/api/v1/inspect`	Inspect a file's structure (sheets, pages, tables)
POST	`/api/v1/read-range`	Bounded read of part of a file
POST	`/api/v1/query-file-table`	Query a table inside a file
POST	`/api/v1/search-file`	Search within a single file

The bounded-read endpoints also have upload variants that accept the file in the request instead of a Storage reference: /api/v1/inspect-upload, /api/v1/read-range-upload, /api/v1/query-file-table-upload, and /api/v1/search-upload.

Collections and search

Method	Endpoint	Description
POST	`/api/v1/collections/index/create`	Create an index from files
PUT	`/api/v1/collections/index/{index_id}/update`	Add or replace files in an index
GET	`/api/v1/collections/index/status/{request_id}`	Indexing request status
POST	`/api/v1/collections/index/{index_id}/cancel`	Cancel indexing
GET	`/api/v1/collections/index/list`	List indexes
GET, DELETE	`/api/v1/collections/index/{index_id}/files`	List or remove indexed files
DELETE	`/api/v1/collections/index/{index_id}`	Delete an index
DELETE	`/api/v1/collections/indexes`	Delete several indexes
POST	`/api/v1/collections/index/metrics`	Size metrics for 1–100 index IDs
POST	`/api/v1/collections/search`	Semantic search
POST	`/api/v1/collections/search/deep`	Start a deep-search job
GET	`/api/v1/collections/search/deep/{job_id}`	Deep-search job result

POST /index/metrics takes index_ids (1–100 unique UUIDs) and returns searchable_chunk_count and index_storage_bytes per index, plus missing_index_ids, measured_index_count, cached_index_count, and failed_indexes.

Knowledge graphs and evaluation

Method	Endpoint	Description
POST	`/api/v1/graph/build`	Build a knowledge graph
GET	`/api/v1/graph/list`	List graphs
GET	`/api/v1/graph/{graph_id}/status`	Graph build status
GET, POST	`/api/v1/graph/{graph_id}/files`	List or add graph files
GET	`/api/v1/graph/{graph_id}/visualization`	Visualization data (`nodes` 1–5000)
GET	`/api/v1/graph/{graph_id}/communities`	Communities (1–500; values outside the range return `422`)
GET	`/api/v1/graph/{graph_id}/suggested-entity-types`	Suggested entity types for a graph
POST	`/api/v1/graph/{graph_id}/approve-entity-types`	Approve entity types and continue the build
POST	`/api/v1/graph/{graph_id}/cancel`	Cancel a graph build
DELETE	`/api/v1/graph/{graph_id}`	Delete a graph
POST	`/api/v1/evaluation/run`, `/api/v1/evaluation/compare`	Run or compare retrieval evaluations
GET	`/api/v1/evaluation/list`, `/api/v1/evaluation/{evaluation_id}`	Evaluation results
GET	`/api/v1/evaluation/{evaluation_id}/questions/{question_id}/chunks`	Retrieved chunks for one question
DELETE	`/api/v1/evaluation/{evaluation_id}/questions`	Remove questions from an evaluation
POST	`/api/v1/evaluation/{evaluation_id}/recompute`	Recompute evaluation scores
GET	`/health/`	Health check

Multi-Tenant and Project Isolation

Indexes are isolated per organization. Search requests can be project-scoped with X-Project-Id; RAGflow then checks cosmos.kb.read and re-authorizes each target index. Deep-search jobs are bound to the subject and project that started them.

Configuration

Variable	Description
`POSTGRES_HOST`	PostgreSQL host
`POSTGRES_USER`	Database user
`POSTGRESPASSWORD`	Base64-encoded password
`POSTGRES_DB`	Database name (`ragflow_dev`)
`EMBEDDING_PROVIDER`	Embedding provider (default `azure`)
`EMBEDDING_MODEL`	Embedding model (default `text-embedding-3-small`)
`GRAPH_STORE_TYPE`	Graph store (default `neo4j`)
`GRAPH_TASK_SOFT_TIME_LIMIT`	Soft time limit for graph tasks (default 4 hours)
`MAX_UPLOAD_BYTES`	Upload limit for the upload endpoints (default 150 MiB)
`INDEXING_HARD_CEILING_BYTES`	Largest file accepted for indexing (default 3 GiB; `0` disables). Excel files use `INDEXING_EXCEL_CEILING_BYTES` (default 100 MB).

Prerequisites

PostgreSQL 15+ with pgvector extension enabled
Redis server for Celery task queue
Tesseract OCR for image document parsing (installed in the image)

bash

# Enable pgvector (run as PostgreSQL admin)
CREATE EXTENSION IF NOT EXISTS vector;

Integration Points

Service	Integration
Configurator	Collection management — creates/deletes indexes via Configurator
Storage	Source files for parsing and indexing, with DMS file authorization
ADK	Agent knowledge base search during conversations
MCP	Inbuilt KB search tool

ℹ️ Info: RAGflow is an optional service. Start it with ./manage.sh dev start knowledge.