Nicolas Cantu
a6bd0ea14c
**Motivations:** - Single operational reference for deploy script vs manual steps and troubleshooting. **Root causes:** - README mixed http-maps manual path with script using conf.d without full operator context. **Correctifs:** - Align documentation with deploy script paths and prerequisites. **Evolutions:** - Expanded README-ia-enso.md (tables, SSRF context, env vars, rotation, troubleshooting). - Feature doc table and deployment pointers; links from docs/README, infrastructure, services. **Pages affectées:** - deploy/nginx/README-ia-enso.md - docs/features/ia-enso-nginx-proxy-ollama-anythingllm.md - docs/README.md - docs/infrastructure.md - docs/services.md
44 lines
2.8 KiB
Markdown
44 lines
2.8 KiB
Markdown
# Feature: Reverse proxy ia.enso.4nkweb.com for Ollama and AnythingLLM
|
||
|
||
**Author:** 4NK team
|
||
|
||
## Objective
|
||
|
||
Expose Ollama and AnythingLLM on the public proxy hostname with HTTPS, path prefixes `/ollama` and `/anythingllm`, and **gate Ollama** with a **Bearer token** checked at the proxy (compatible with Cursor’s OpenAI base URL + API key).
|
||
|
||
## Impacts
|
||
|
||
- **Proxy (nginx):** new `server_name`, TLS, locations, HTTP `map` for Bearer validation; maps deployed under `/etc/nginx/conf.d/` when using the provided script.
|
||
- **Backend (192.168.1.164):** must accept connections from the proxy on `11434` and `3001`; Ollama must not rely on the client `Authorization` header (nginx clears it after validation).
|
||
- **Clients:** Cursor uses `https://ia.enso.4nkweb.com/ollama/v1` and the shared secret as API key; avoids private-IP SSRF blocks in Cursor when the hostname resolves publicly from the client infrastructure.
|
||
|
||
## Repository layout
|
||
|
||
| Path | Purpose |
|
||
|------|---------|
|
||
| `deploy/nginx/sites/ia.enso.4nkweb.com.conf` | `server` blocks, upstreams to `192.168.1.164` |
|
||
| `deploy/nginx/http-maps/ia-enso-ollama-bearer.map.conf.example` | Example Bearer `map` (manual install) |
|
||
| `deploy/nginx/http-maps/websocket-connection.map.conf.example` | Example WebSocket `map` (manual install) |
|
||
| `deploy/nginx/deploy-ia-enso-to-proxy.sh` | SSH deploy: maps + site, `nginx -t`, reload; Bearer-only retry if websocket `map` already exists |
|
||
| `deploy/nginx/README-ia-enso.md` | **Operator reference:** automated + manual steps, env vars, checks, troubleshooting |
|
||
|
||
## Deployment modalities
|
||
|
||
**Preferred:** run `./deploy/nginx/deploy-ia-enso-to-proxy.sh` from `smart_ide` on a host with SSH access (see `README-ia-enso.md` for prerequisites and environment variables).
|
||
|
||
**Manual:** DNS → TLS (certbot) → install `map` directives inside `http { }` (via `conf.d` or `http-maps` includes) → install site under `sites-available` / `sites-enabled` → `nginx -t` → reload. Details: `deploy/nginx/README-ia-enso.md`.
|
||
|
||
Restrict backend ports on `192.168.1.164` to the proxy source where a host firewall is used.
|
||
|
||
## Analysis modalities
|
||
|
||
- `curl` to `/ollama/v1/models` with and without `Authorization: Bearer <secret>` (expect 200 / 401).
|
||
- Browser access to `/anythingllm/` and application login.
|
||
- Cursor connectivity after configuration (no `ssrf_blocked` if the hostname does not resolve to a blocked private IP from Cursor’s perspective).
|
||
|
||
## Security notes
|
||
|
||
- The Bearer secret is equivalent to an API key; rotate by updating the `map` file and client configs together.
|
||
- AnythingLLM remains protected by **its own** application authentication; the `/anythingllm` location does not add the Ollama Bearer gate.
|
||
- A public URL for `/ollama` exposes the inference endpoint to anyone who knows the secret; combine with network controls if required.
|