0 introduction

0. Introduction: What is tripleF?

tripleF (3F) — short for fkoff — is an open-source agentic AI workbench. It is a direct answer to the proprietary chat platforms: same ambition, no lock-in, no meter running on your own hardware. Every capability is built on free, open-source models that run fully locally — or, when you want more headroom, on Ollama Cloud models. Your conversations, images, and tooling stay under your control either way.

tripleF is early in development and unapologetic about aiming for the top of the open-source community. It already ships a complete chat experience — including features some proprietary chats still miss — and it is designed as a platform, not a demo.

What 3F does today

• Full chat experience — multi-image conversations, streaming answers rendered live, reasoning/thinking areas that show the model's chain of thought, session-scoped conversation history, and conversation compaction that compresses long threads into a context-efficient summary without losing the plot.
• Local-first inference — every request is served by Ollama. Point OLLAMA_HOST at your own machine for fully-offline operation, or set OLLAMA_API_KEY to use Ollama Cloud models. The workbench treats both as one continuum.
• Agentic harness — requests flow through a deterministic step engine: sanitize → interpret → execute → respond. The harness classifies intent, selects tools (web/image/news/shopping/places/business-reviews/video search via Serper, YouTube Data API video search, webpage scrape/fetch), language-detects results and files foreign-language finds into an international-coverage aside, enforces structured output schemas, and validates responses before they reach the UI.
• Structured vision & media answers — describe, compare, OCR, imagine, news, article, product, image/video list schemas turn image understanding into machine-readable, UI-renderable results instead of plain text blobs.
• Real-time by construction — answers stream token-by-token over Socket.IO rooms; a request can be cancelled mid-flight by the user, and the worker honours the cancellation token at step boundaries.
• Operability built in — BullMQ queues with retry/backoff, a persisted dead-letter queue (replay, edit, re-instate), queue and system health consoles, provider-override management, and image preprocessing controls — all exposed in the dashboard's SysCtl area.

Roadmap

The trajectory is deliberately aggressive:

• Speech-to-text input
• Music and image generation alongside understanding
• Provider integrations — YouTube, messengers, finance, and other everyday data sources
• Anything the proprietary incumbents ship — shipped here, open, first.

System Overview

                      ┌──────────────────────────────────────────────┐
                      │                DASHBOARD (Vue 3)             │
                      │  Chat · SysCtl · DLQ · PProc · Themes        │
                      └───────────────▲───────────────┬──────────────┘
                        Socket.IO     │               │ REST (Fastify,
                        (rooms)       │               │ /api/v1, Swagger)
                      ┌───────────────┴───────────────▼──────────────┐
                      │              SERVER (NestJS + Fastify)       │
                      │                                              │
                      │  Harness controller ──▶ BullMQ harness queue │
                      │                              │               │
                      │                     Harness processor        │
                      │              step engine: sanitize →         │
                      │      interpret → execute → respond           │
                      │         │              │          │          │
                      │  Ollama/AI SDK   Serper/YouTube Schemas      │
                      └─────┬──────────┬──────────┬──────────┬───────┘
                            │          │          │          │
                     ┌──────▼───┐ ┌────▼─────┐ ┌──▼───────┐ ┌▼────────┐
                     │ KeyDB    │ │ Postgres │ │  MinIO   │ │ Ollama  │
                     │ (queues, │ │ (Prisma: │ │ (image   │ │ (local  │
                     │ sockets) │ │ convos,  │ │ payloads)│ │ /cloud) │
                     │          │ │ DLQ,     │ │          │ │         │
                     │          │ │ config)  │ │          │ │         │
                     └──────────┘ └──────────┘ └──────────┘ └─────────┘

Technology Matrix

Layer	Technology	Role
API runtime	NestJS 11 + Fastify 5	HTTP, versioning, multipart intake, Swagger
Agent stack	Vercel AI SDK 6 + Ollama	Model calls, tool loop, streaming
Inference	Ollama (local or Ollama Cloud)	Vision-LLM inference + embeddings
Async	BullMQ + KeyDB (Redis-compatible)	Job queue, retries, cancellation tokens, socket pub/sub
Persistence	PostgreSQL 16 + Prisma 7	Conversations, dead-letter records, system config
Object storage	MinIO (S3 API)	Image payloads per session/conversation
Real-time	Socket.IO 4	Result streaming rooms, cancellation, events
Frontend	Vue 3 + Vite 8, Pinia, TanStack Query, Tailwind v4	Workbench UI
Tooling	pnpm workspaces, TypeScript, Vitest, ESLint, Docker Compose	Monorepo ergonomics

Repository Layout

triplef.io/
├── server/                  # 3F server (NestJS + Fastify)
│   ├── src/modules/         # harness, ai-sdk, bullmq, dead-letter, minio,
│   │                        # persistence, sharp, socket-io, provider-overrides ...
│   ├── prisma/              # schema: HarnessConversation, HarnessDlq, HarnessConfig
│   └── Dockerfile           # standalone image (kept for a potential repo split)
├── dashboard/               # 3F dashboard (Vue 3 + Vite)
│   ├── src/components/      # chat, sysctl, dlq, pproc, app shell, widgets
│   └── Dockerfile           # standalone image (kept for a potential repo split)
├── Dockerfile               # canonical monorepo image (all targets)
├── compose.yml              # dev stack: deps + server + dashboard
├── infra.compose.yml        # postgres, keydb, minio (ollama/searxng optional)
└── .wiki/                   # this documentation

License & Ownership

MIT-licensed. Built with AI-assisted context coding — every architectural decision reviewed and owned by humans.