Which platform runs on a small VPS?
Short answer
anythingllm — AnythingLLM, without a trade this time. It holds 542 MB in one container and also had the highest retrieval accuracy of anything measured on this corpus.
What this use case actually needs
- The host is a small VPS — two to four gigabytes of RAM, likely shared with other services
- Container count matters as much as memory, because each one carries supervision and restart cost
- Documents are ordinary internal files rather than large filings
How the candidates measured up
| Platform | Documents ingested | Parsing | Retrieval | Peak memory | Ready |
|---|---|---|---|---|---|
| anythingllm | 36/36 | 98% | 83% | 542.2 MB | 146 s |
| openwebui | 36/36 | 46% | 73% | 980.7 MB | 18 s |
| dify | 36/36 | 100% | 68% | 2,619.7 MB | 39 s |
| fastgpt | 36/36 | 100% | 74% | 3,478.8 MB | 239 s |
Every figure comes from the same run against the same documents, so these are directly comparable. The full per-document and per-query rows are in the raw data.
The trade that used to exist here has gone
On our first corpus this page had to hedge: AnythingLLM won on footprint but ingested six documents out of eleven, so choosing it meant accepting an incomplete knowledge base. On a corpus of ordinary internal documents it ingested all thirty-six and scored highest on retrieval as well.
So on this workload it is simply the recommendation, on both dimensions at once.
| Platform | Containers | Peak memory | Ingested | Retrieval |
|---|---|---|---|---|
| AnythingLLM | 1 | 542 MB | 36/36 | 83% |
| Open WebUI | 1 | 981 MB | 36/36 | 73% |
| Dify | 15 | 2,620 MB | 36/36 | 68% |
| FastGPT | 13 | 3,479 MB | 36/36 | 74% |
| RAGFlow | 5 | 9,960 MB | 36/36 | 73% |
The spread is eighteen-fold, and container count does not predict it
RAGFlow runs five containers and needs 9,960 MB. FastGPT runs thirteen and needs 3,479 MB. Dify runs fifteen and needs 2,620 MB — the most containers and the third-lowest memory.
The reason is what is inside them: RAGFlow bundles Elasticsearch, which holds several gigabytes before a document arrives. A requirements page quoting either number alone tells you very little about whether a platform fits your host.
Open WebUI is the alternative worth weighing
981 MB in one container, ready to serve in eighteen seconds — the fastest start of anything measured — and 73% retrieval. It costs roughly twice AnythingLLM’s memory for ten points less accuracy.
Its weakness is parsing: 46%, by far the lowest observed. That figure is approximate, since it exposes no chunk listing and can only under-report, but the gap to the others is too wide to be entirely an artefact. If your documents carry precise values you need back intact, this is the wrong end of the trade.
What a small host cannot have
None of the platforms that expose exact chunk visibility — FastGPT, Dify, RAGFlow — fit comfortably in under two gigabytes. If auditability of stored text is a requirement, a small VPS is not the right host, and that is a real constraint rather than a preference.
Where this recommendation stops applying
- Your documents are megabyte-scale; AnythingLLM embeds each one synchronously during upload and did not finish half of a large-document corpus
- You need to audit stored chunks, which it does not expose
- The port is reachable before you add authentication in front of it