Ollama HTTP Endpoints

All examples assume Ollama is running locally and reachable at http://127.0.0.1:11434. PAI starts Ollama via ollama.service. See reference.md for the loopback posture and the known OLLAMA_HOST=0.0.0.0 caveat.

Warning — exposing Ollama beyond loopback. These endpoints have no authentication. Do not bind to a public interface without adding your own auth layer (reverse proxy, SSH tunnel, VPN).

Schemas below are summaries. The upstream Ollama API docs are the source of truth; if you find a discrepancy, trust upstream.

---

POST /api/generate

Single-turn completion from a prompt.

Request body

{
  "model": "phi3:mini",
  "prompt": "Why is the sky blue?",
  "stream": true,
  "options": { "temperature": 0.7 }
}

Response

NDJSON stream (one JSON object per line) when stream: true (default). Each chunk has a response field with a token fragment and done: false, terminated by a final object with done: true and summary fields (total_duration, eval_count, etc.). With stream: false, a single JSON object is returned.

Errors

• 404 — unknown model. Pull it first (/api/pull or ollama pull).
• 400 — malformed body.
• 500 — runtime error (out of memory, GPU failure).

Example

curl http://127.0.0.1:11434/api/generate \
  -d '{"model":"phi3:mini","prompt":"Say hi","stream":false}'

---

POST /api/chat

Multi-turn chat completion. Preferred over /api/generate for conversational use because it preserves role structure.

Request body

{
  "model": "phi3:mini",
  "messages": [
    { "role": "system", "content": "You are terse." },
    { "role": "user", "content": "Hello." }
  ],
  "stream": true
}

Response

NDJSON stream. Each chunk contains a message object (role, content) and done. Final chunk has done: true plus duration/token counts.

Errors

• Same as /api/generate.

Example

curl http://127.0.0.1:11434/api/chat \
  -d '{"model":"phi3:mini","messages":[{"role":"user","content":"hi"}],"stream":false}'

---

GET /api/tags

List locally installed models.

Response

{
  "models": [
    {
      "name": "phi3:mini",
      "modified_at": "2026-03-01T12:00:00Z",
      "size": 2147483648,
      "digest": "sha256:…"
    }
  ]
}

Example

curl http://127.0.0.1:11434/api/tags

---

POST /api/show

Show metadata for a model (modelfile, parameters, template, license).

Request body

{ "name": "phi3:mini" }

Response

JSON object with modelfile, parameters, template, details.

Example

curl http://127.0.0.1:11434/api/show -d '{"name":"phi3:mini"}'

---

POST /api/pull

Download a model from the Ollama registry.

Request body

{ "name": "llama3.2:3b", "stream": true }

Response

NDJSON progress stream (status, digest, total, completed). Terminates with { "status": "success" }.

Errors

• 404 — model not in registry.
• Network errors if egress is blocked by UFW. PAI firewall defaults deny outbound; you must allow DNS + HTTPS to the Ollama registry for pulls to succeed. See shared/hooks/live/0500-firewall.hook.chroot.

Example

curl http://127.0.0.1:11434/api/pull -d '{"name":"llama3.2:3b"}'

---

DELETE /api/delete

Remove an installed model.

Request body

{ "name": "llama3.2:3b" }

Response

200 OK on success, 404 if not found.

Example

curl -X DELETE http://127.0.0.1:11434/api/delete \
  -d '{"name":"llama3.2:3b"}'

---

POST /api/embeddings

Compute embedding vectors for a prompt.

Request body

{ "model": "nomic-embed-text", "prompt": "hello world" }

Response

{ "embedding": [0.0123, -0.456, ...] }

Example

curl http://127.0.0.1:11434/api/embeddings \
  -d '{"model":"nomic-embed-text","prompt":"hello"}'

---

GET /api/ps

List models currently loaded in memory.

Response

{
  "models": [
    { "name": "phi3:mini", "size_vram": 2147483648, "expires_at": "…" }
  ]
}

Example

curl http://127.0.0.1:11434/api/ps

---

Streaming semantics

/api/generate, /api/chat, and /api/pull stream NDJSON by default. Each newline-terminated line is one JSON object. A client should parse line-by-line and stop when it sees "done": true (generate/chat) or "status": "success" (pull).

To disable streaming and receive a single JSON response on generate/chat, set "stream": false.

---

Differences from upstream

PAI ships upstream Ollama unchanged. Only how it is launched differs:

Setting	PAI default	Configured in
Bind address	0.0.0.0:11434 (see caveat)	shared/includes/etc/systemd/system/ollama.service
User / group	ollama / ollama	same
Model storage path	/usr/share/ollama/.ollama/models	HOME=/usr/share/ollama in the unit file
Auto-start on boot	Enabled via WantedBy=multi-user.target	same
Firewall	UFW default-deny inbound/outbound	shared/hooks/live/0500-firewall.hook.chroot

No PAI-specific HTTP endpoints are added, removed, or renamed.