RAGFlow 1.0 in Docker: a 361MB Image, Six Containers

📘 Tutorials 2026-11-20 7 min read

The install commands in old blog posts still work, then the stack dies quietly: one container exits, the web port returns 502, and nothing in the log says which of the six dependencies failed.

📜 Table of Contents

RAGFlow tagged v1.0.0 on 6 November 2026. The number that matters for anyone installing it is not a benchmark: the published image went from 3,618 MB at v0.27.3 to 361 MB at v1.0.0. Same repository, same registry, ten times smaller. Both sizes came from the Docker Hub API today, so this is a pull-and-see number, not a changelog claim.

The second change is the container count. A stock docker compose up -d with the default .env starts six containers, and four of them are not RAGFlow. If any of the four refuses to come up, the web interface still serves a login page and then fails on every request.

factvalue (checked 20 Nov 2026)
latest releasev1.0.0, published 2026-11-06
GitHub stars / forks94,440 / 11,351
open issues1,293
licenseApache-2.0
image size, v1.0.0 vs v0.27.3361 MB vs 3,618 MB
documented minimum host4 cores, 16 GB RAM, 50 GB disk
Docker / Compose24.0.0+ / v2.26.1+ with BuildKit

The four commands

Everything below comes from the repository's own docker/README.md and docs/quickstart.mdx at the v1.0.0 tag. Images exist for x86-64 and ARM64 from this release on.

# 1. Elasticsearch needs a bigger mmap count than most distros allow
sysctl vm.max_map_count
sudo sysctl -w vm.max_map_count=262144
# then add vm.max_map_count=262144 to /etc/sysctl.conf by hand,
# or the value is gone at the next reboot

# 2. clone and pin the tag
git clone https://github.com/infiniflow/ragflow.git
cd ragflow/docker
git checkout v1.0.0

# 3. start it
docker compose -f docker-compose.yml up -d

# 4. health, before you open a browser
curl -f http://localhost/api/v1/system/healthz

Step 4 is the one people skip. The endpoint walks the API and its dependencies and returns 200 only after the database migrations finish. First boot runs migrations inside the RAGFlow container, so docker ps can show everything Up while the health check still returns 000.

What actually runs

The default .env resolves to COMPOSE_PROFILES=elasticsearch,cpu,metadata-mysql. Three services carry no profile at all, so they start regardless: minio, kvrocks, nats. With MySQL and Elasticsearch that makes six containers holding the stack together.

containerimagesizewhy it is there
ragflow-cpuinfiniflow/ragflow:v1.0.0361 MBAPI server, ingestion workers, migrations
es01elasticsearch:8.11.3673 MBdocument engine, chunks and vectors
mysqlmysql:8.0.40171 MBmetadata: users, datasets, task state
miniopgsty/silo, MinIO-compatible55 MBoriginal file storage
kvrocksapache/kvrocks:2.16.065 MBcache and queue for the Go services
natsnats:2.14.27 MBJetStream bus for ingestion and sync

Sum the images: 1,332 MB. The whole stack pulls smaller than the single 3.6 GB RAGFlow image used to, which matters on a 40 GB cloud disk where images, volumes and Elasticsearch indices share the same space.

Four settings that decide whether it boots

1. vm.max_map_count, only with Elasticsearch

Elasticsearch maps index segments into memory and will not start below 262144. The repo notes this is usually unnecessary with Infinity, the other bundled engine. sysctl -w sets a temporary value, and that is how a stack that ran fine for a week stops starting after a kernel update.

2. Port 80

SVR_WEB_HTTP_PORT=80 is the default. Any host already running nginx or Caddy on 80 gets bind: address already in use, and compose leaves the rest of the stack up. Change it in docker/.env to SVR_WEB_HTTP_PORT=8080, then the health check and the browser URL both need the port.

3. The 7.52 GiB Elasticsearch cap

MEM_LIMIT=8073741824 is 7.52 GiB, applied as mem_limit to es01. On a 16 GB host that leaves room for MySQL, the RAGFlow workers and the OS. On an 8 GB box the container is capped under what the JVM wants, and you get restart loops that look like a crash rather than an out-of-memory kill. The documented 16 GB floor exists for this reason.

4. DOC_ENGINE and the profile it implies

DOC_ENGINE is not a hint, it selects containers through COMPOSE_PROFILES=${DOC_ENGINE},${DEVICE},metadata-${METADATA_DB_PROFILE}. Engines defined in the base compose file: elasticsearch (default), infinity, opensearch, oceanbase, seekdb, serenedb, vastbase. Switching to infinity starts a 325 MB container instead of the 673 MB Elasticsearch one and drops the mmap requirement, but chunks already written to Elasticsearch do not migrate themselves.

Ports after 1.0

settingdefaultwhat is on it
SVR_WEB_HTTP_PORT80web UI, proxied
SVR_HTTP_PORT9380API, including POST /api/v1/mcp
ADMIN_SVR_HTTP_PORT9381admin server, switched on by the compose flags
ES_PORT1200Elasticsearch, host side
EXPOSE_MYSQL_PORT3306MySQL
MINIO_PORT / console9000 / 9001object storage
KVROCKS_PORT6379kvrocks
SVR_MCP_PORT9382optional MCP listener, commented out in compose

Elasticsearch is on 1200, not 9200. Firewall rules and monitoring probes written for a 0.27 deployment now point at nothing.

Ship it with different passwords

The repository's quickstart says to change the default credentials before putting this on a network-accessible host. They are all infini_rag_flow: Elasticsearch, MySQL, MinIO, kvrocks. Two more defaults are worth knowing: ADMIN_DEFAULT_PASSWORD=admin for the admin server on 9381, and REGISTER_ENABLED=1, which lets anyone who reaches port 80 create an account.

The passwords come from MINIO_PASSWORD, MYSQL_PASSWORD, ELASTIC_PASSWORD and the kvrocks-side REDIS_PASSWORD, so you edit one file and restart. Changing them once data exists means recreating the volumes.

What 1.0 rewrote

Useful before you file a bug against a 0.27 guide. The API layer is Go now and serves POST /api/v1/mcp on 9380; ingestion, sync and knowledge-compilation messages ride NATS JetStream; and in the open-source 1.0 build, DeepDoc layout analysis, OCR and table recognition run on CPU. That last point sets the throughput expectation: CPU-only parsing of a 50-page PDF is slow by design, and a GPU does not change it in this build.

Upgrading an existing deployment has its own trap. If a checkout's code version is older than the database migration version, the server refuses to start with Refusing to start: database was migrated by a newer version. For local development the flag is RAGFLOW_DEV_MODE=true; in production the fix is a build at least as new as the migration, not the flag.

When it does not come up

symptomcausecheck
web UI 502, containers UpAPI still migratingcurl -f localhost/api/v1/system/healthz, then docker logs -f ragflow-cpu
es01 restarts every minutemmap count below 262144sysctl vm.max_map_count
compose fails on port 80nginx already boundset SVR_WEB_HTTP_PORT, add the port to every URL
es01 killed under loadthe 7.52 GiB limitdocker stats; raise MEM_LIMIT only with host RAM to match
old monitoring silentES moved to port 1200repoint probes from 9200

FAQ

Does the 361 MB image mean 1.0 dropped features? No. The image shrank because the Python bundle was split and the Go API took over the request path; parsers, document engines and all six containers are still there.

Can I run it on 8 GB of RAM? The documented floor is 16 GB with a 7.52 GiB Elasticsearch cap. Below that, either switch DOC_ENGINE to infinity and lower MEM_LIMIT, or expect restarts under parse load.

Where do I configure models? After the first login, in the model provider settings. Any OpenAI-compatible endpoint works, and embedding and rerank models are separate entries from the chat model.

Is there an API key system? Yes, per-user keys issued in the UI, used against port 9380. The MCP endpoint sits on the same port.

Numbers here were pulled on 20 November 2026 from the Docker Hub API (image sizes), the GitHub API (stars, forks, issues, release date) and the v1.0.0 tag of the repository (docker/.env, docker-compose-base.yml, docs/quickstart.mdx). Star counts and image tags move, so re-check before quoting.

Related: what a thousand RAG questions costs, RAGFlow against Dify.

Related Articles
2026-07-22
Cursor AI Alternative 2026: 8 Free IDE Assistants
2026-08-13
Run LLM on Android 2026: 5 Ways to Get a Local Model on Your Phone
2026-07-19
Open Source TTS: Free Text-to-Speech Solutions
2026-08-03
Telegram AI Bot 2026: python-telegram-bot (29k Stars) + LLM - Free, Official, and Easy
2026-08-11
Synthetic Data Generation 2026: Tools and Methods for Training Better Models
2026-07-16
AI Quantization Techniques 2026

Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.

💬 Comments (0)

No comments yet. Be the first!

Login to comment