Vibe Content Lab

Portfolio-grade internal AI tool for a public API documentation corpus. It demonstrates a complete applied-AI delivery path: RAG, embeddings, local vector search, retrieval evaluation, FastAPI, React, n8n, Dify, Docker, logs, webhooks and human approval before paid actions.

Honest scope

This is an independent technical lab, not a product of VibeMarketolog and not a claim of affiliation. The knowledge base is built only from its public Agent API documentation. The application does not issue a paid generation until an operator reviews a free /generate/estimate result and explicitly confirms it.

Live deployment

• Demo: lab.syntog.ru, protected with Basic Auth.
• Operator automation UI: n8n.lab.syntog.ru, protected for operator access.

The deployment demonstrates Docker, nginx reverse proxying, TLS, Basic Auth and loopback-only application ports. Provider credentials are intentionally absent until a separate project key and spending limits are configured.

For a qualification-focused evidence map, see docs/PORTFOLIO_EVIDENCE_RU.md.

Components

Component	Evidence of skill
rag/	document fetch, chunking, OpenAI text-embedding-3-small, persistent local Chroma, source citations
rag/evaluation/	ground-truth retrieval set, hit@k and MRR metrics
backend/	FastAPI, SQLite audit trail, Vibe API adapter, HMAC webhook validation, explicit cost approval
frontend/	React/Vite internal UI for search, estimate, confirmation and feedback
n8n/	importable scheduled health-and-digest workflow, Telegram delivery
dify/	Dify Cloud instructions, citation-first system prompt and acceptance checks
deploy/	Docker Compose, nginx templates, resource limits and operational runbook

Security controls

• Secrets are server-only in /opt/vibe-content-lab/.env, never in Git or browser JavaScript.
• The backend calls /generate/estimate first. POST /generations/confirm needs the returned one-time estimate ID.
• The Vibe API key should have only read and generate scopes, a small daily_spend_limit, and an IP allowlist for the server.
• Incoming generation webhooks are validated with X-Vibe-Signature and a constant-time HMAC comparison.
• n8n.lab.syntog.ru is a private operator interface behind nginx Basic Auth.
• The Dify component uses the separate private Dify Cloud workspace, rather than copying Factory credentials or trying to run Dify below its vendor memory minimum.

Setup

Development harness

On Windows, all routine work is behind one command:

.\scripts\harness.ps1 doctor     # tool and non-secret configuration status
.\scripts\harness.ps1 bootstrap  # isolated Python environment + locked UI dependencies
.\scripts\harness.ps1 check      # compile, Ruff, JSON validation, React production build
.\scripts\harness.ps1 test       # unit tests, no API keys and no paid requests
.\scripts\harness.ps1 api        # FastAPI on localhost:8000
.\scripts\harness.ps1 ui         # Vite UI on localhost:5173
.\scripts\harness.ps1 index      # fetch public docs and build Chroma index
.\scripts\harness.ps1 eval       # evaluate retrieval against ground truth

bootstrap, check, and test work without API keys. index and the LLM answer endpoint require a separate OpenAI key in local .env; paid Vibe requests additionally require an explicit estimate and confirmation.

Server setup

1. Create /opt/vibe-content-lab/.env on the server using .env.example, then chmod 600 it.
2. Set a separate OpenAI project key and a Vibe API key. Do not reuse Factory credentials.
3. Start with docker compose --env-file /opt/vibe-content-lab/.env up -d --build.
4. Build the knowledge index: docker compose exec api python -m rag.ingest.
5. Run retrieval evaluation: docker compose exec api python -m rag.evaluation.run.

The deploy and workflow steps are intentionally not claimed as completed until they have real execution records.