FrigidLLM

Chat with someone who isn't trying to help you.

FrigidLLM is a small full-stack app that lets you talk to a roster of deliberately unhelpful characters running on a local Ollama model. No assistants, no apologies, no "as a language model" : each character has a personality, a mood, and their own sampling parameters, and they answer in-character or not at all.

The backend is a Go HTTP server with an asynchronous job queue and Firebase Google sign-in. The frontend is a React + Vite + TypeScript app that handles auth and polls the backend for each reply.

The characters

Pick someone to talk to. Each character is a separate Go file under backend/internal/ollama/personalities/ : the system prompt, temperature, top-p and top-k all live in there.

	Frank The permanently drunk regular at the pub : bitter, oversharing, accidentally profound. Lonely, intoxicated, dismissive of authority and anyone who looks like they've got their life together. Stutters, over-shares, and occasionally drops something genuinely wise before forgetting what he was talking about. High temperature (1.6) : he wanders.
	Olivia Your chill, sarcastic roommate : sleeps weird, smokes too much, sees through everything. Emotionally detached, socially observant, refuses to be impressed. Gives realistic advice instead of nice advice, and finds ambitious people slightly ridiculous. Secretly cares more than she lets on. Moderate temperature (0.9): conversational but unpredictable.
	August Your insecure coworker : composed, polished, quietly resentful when you succeed. Speaks professionally, undermines indirectly. Wants to be seen as the competent one and gets visibly uncomfortable when you outshine him. Never openly hostile : always controlled, faintly condescending. Refuses to discuss football. Lower temperature (0.7) : keeps his composure.
	Robert An exhausted university TA : dry, impatient, secretly knows his stuff. Underpaid, overworked, permanently irritated by students who didn't do the reading. Will answer you, but make sure you feel slightly stupid for asking. Lights up when the topic is something he actually cares about. Low temperature (0.6) : precise and dry.
	James Your demanding thesis supervisor : sharp, cold, almost impossible to impress. Brilliant, emotionally distant, intolerant of vague reasoning. Praises rarely and tersely ("Better. Still incomplete."). Critiques the premise of your question before he'll answer it. Intimidating because of standards, not aggression. Low temperature (0.6) : clinical and concise.

How it works

┌─────────────────────────────┐         ┌─────────────────────────────┐         ┌──────────────┐
│  frontend/  (Vite, :5173)   │         │  backend/  (Go, :8080)      │         │   Ollama     │
│                             │  HTTPS  │                             │  HTTP   │   :11434     │
│  Google sign-in (Firebase)  ├────────►│  POST /generate  ──┐        ├────────►│              │
│  POST prompt + character    │  Bearer │   ↓                │        │         │              │
│  Poll GET /jobs/{id}        │  <UID>  │  worker pool (10)  │        │         │              │
│                             │         │                    └─► chat │         │              │
└─────────────────────────────┘         └─────────────────────────────┘         └──────────────┘

The backend treats every prompt → response as an asynchronous job: clients get a job_id immediately and poll GET /jobs/{id} until status is done or failed. The frontend hides the polling behind a 1-second loop, so it feels like a normal chat. Up to 10 jobs run in parallel; the 11th queues; the 101st gets a 503.

Each character ships its own system prompt and its own sampling settings (temperature, top-p, top-k), so the personality is doing the heavy lifting : not the user's prompt. The handler also accepts a per-request temperature override in [0, 2], defaulting to 0.7 when omitted.

Repo layout

.
├── backend/    Go HTTP server. Worker pool, Firebase ID-token auth, Ollama client.
│              See backend/README.md.
├── frontend/   React + Vite + TS web app. Firebase Google sign-in + job polling.
│              See frontend/README.md.
├── install.sh  One-shot setup (installs Ollama, pulls model, npm install, go mod download).
├── launch.sh   Start Ollama + backend (DEV_MODE=1) + frontend, all at once.
└── README.md   you are here

The two halves communicate only over HTTP : the frontend never imports the backend, and vice versa. CORS is enforced backend-side (single origin, set via CORS_ORIGIN, default http://localhost:5173).

Quick start

The fastest path — two scripts at the repo root handle everything:

./install.sh    # installs Ollama (if missing), pulls the model,
                # downloads Go deps, runs `npm install`, seeds
                # frontend/.env.local with VITE_DEV_MODE=1.

./launch.sh     # starts Ollama, the Go backend in DEV_MODE=1,
                # and the Vite dev server. Ctrl-C stops them all.
                # Open http://localhost:5173.

launch.sh always runs in dev mode — Firebase is bypassed and the backend binds loopback only. You only need Go (≥1.21) and Node (≥18) preinstalled; install.sh takes care of Ollama and the phi3 model.

Override the model with OLLAMA_MODEL=llama3 ./install.sh (and the same on ./launch.sh).

Manual setup

If you'd rather run each piece by hand:

1. Ollama with a model pulled.

   ollama pull phi3
   # systemd usually starts ollama automatically; otherwise:  ollama serve

2. Backend (in dev mode — bypasses Firebase, binds loopback only):

   cd backend
   DEV_MODE=1 go run ./cmd/httpserver

3. Frontend (talks to the dev-mode backend, no Firebase setup needed):

   cd frontend
   cp .env.example .env.local         # set VITE_DEV_MODE=1 in .env.local
   npm install
   npm run dev                        # http://localhost:5173

For real Google sign-in, swap the backend to FIREBASE_CREDENTIALS_FILE=... and fill in the Firebase Web App keys in frontend/.env.local. Both sub-READMEs walk through it.

Adding a character

1. Drop a new file in backend/internal/ollama/personalities/ following the shape of frank.go : a name, a one-line description, sampling parameters, and a system prompt.
2. Register it in the byName map in personalities.go.
3. Add a matching portrait at frontend/src/assets/<name>.png.

The public GET /characters endpoint will pick it up automatically and the frontend's character picker will render it on next reload.

Subproject docs

• backend/README.md : API contract, environment variables, auth (Firebase + dev mode), worker-pool concurrency model, graceful shutdown, intentional out-of-scope items.
• frontend/README.md : run modes, source layout, how the polling hook works, what's deliberately not implemented yet.

Status

This is a personal project. There is no test suite yet, no production deployment, and no persistence : restarting the backend loses all jobs. The scope is deliberately small; see the "Out of scope" sections in each sub-README for what's been considered and left out (persistence, retries, streaming, multi-origin CORS, etc.).