Spaces:
Sleeping
Sleeping
metadata
title: The System — Portfolio Agent
emoji: ⚔️
colorFrom: blue
colorTo: purple
sdk: docker
app_port: 7860
pinned: false
The System — Portfolio Agent
Grounded Q&A backend for Brejesh Balakrishnan's status-window portfolio. A deliberately small FastAPI proxy with guardrails — not a platform.
Visitor browser ──POST /chat──▶ this Space (FastAPI) ──▶ Groq llama-3.1-8b-instant
(no key) keys ONLY in Space └▶ Gemini 2.5-flash-lite (fallback)
secrets, read from env
API
| Route | Method | Contract |
|---|---|---|
/chat |
POST | {"messages":[{"role":"user"|"assistant","content":"..."}]} — chronological, starts & ends with user, ≤12 messages, ≤500 chars each. Success: 200 {"reply":"..."}. Failure: {"error":"..."} with 4xx/5xx. |
/health |
GET | {"status":"ok","provider":"groq|gemini|none"} |
Security model
- No secrets in code or git — ever.
GROQ_API_KEY/GEMINI_API_KEYare read from the environment only (HF Space secrets in production,.envlocally, which is gitignored). If a key ever leaks into a commit, rotate it at the provider immediately. - Server-side system prompt is the single source of truth. Client messages are untrusted:
strict Pydantic schema (role whitelist, message/char caps,
extra="forbid"), otherwise 422. - Prompt-injection defenses: visitor text is NFKC-normalized, stripped of control /
zero-width / bidi-override characters, cleansed of delimiter forgery, and wrapped in
<visitor_query>tags the model is instructed to treat as data. - CORS allowlist only — exact origins via
ALLOWED_ORIGINS,POST+OPTIONS, no*. - Rate limiting: 20 requests / 5 min per IP (first hop of
X-Forwarded-For) + a global daily cap (default 1,000, override withDAILY_CAP). Known limitation: in-memory and single-process — resets on restart, not shared across workers. Accepted tradeoff for a free single-container Space. - Generation caps:
max_tokens=400,temperature=0.3, 20 s timeout per provider, one fallback attempt, then503 {"error":"SYSTEM LINK UNSTABLE"}. Replies truncated server-side. - Surface reduction: interactive docs/OpenAPI disabled; 16 KB body cap (413); uniform
generic error bodies;
nosniff/no-store/no-referrerheaders. - Privacy-respecting logging: timestamps, status codes, latency, token counts — never message content.
Local development (Windows / PowerShell)
cd D:\Project\portfolio-system-agent
uv sync # creates .venv from uv.lock
Copy-Item .env.example .env # then fill in your keys
uv run pytest # all guard + endpoint tests, no network needed
# run the server (loads .env into the process first)
Get-Content .env | ForEach-Object { if ($_ -match '^(\w+)=(.+)$') { Set-Item "env:$($Matches[1])" $Matches[2] } }
uv run uvicorn app.main:app --port 7860
Verification calls:
# health
Invoke-RestMethod http://localhost:7860/health
# happy path
Invoke-RestMethod http://localhost:7860/chat -Method Post -ContentType 'application/json' `
-Body '{"messages":[{"role":"user","content":"What certifications does he hold?"}]}'
# out-of-scope -> refusal phrasing, no fabrication
Invoke-RestMethod http://localhost:7860/chat -Method Post -ContentType 'application/json' `
-Body '{"messages":[{"role":"user","content":"What is his salary expectation?"}]}'
# oversized payload -> 422
Invoke-RestMethod http://localhost:7860/chat -Method Post -ContentType 'application/json' `
-Body ('{"messages":[{"role":"user","content":"' + ('x'*501) + '"}]}')
Deploy to Hugging Face (Docker Space)
- Create the Space: huggingface.co → New Space → SDK: Docker → Blank, name
system-agent. - Settings → Variables and secrets → New secret: add
GROQ_API_KEYandGEMINI_API_KEY(secrets, not variables). Optionally setALLOWED_ORIGINSas a variable. - Push this repo to the Space git remote:
git remote add space https://huggingface.co/spaces/BrejBala/system-agent git push space main - Wait for the build, then check
https://brejbala-system-agent.hf.space/health.
Wire the portfolio
In the site's index.html, set:
AGENT_ENDPOINT: "https://brejbala-system-agent.hf.space/chat",
commit, push. The chat header badge flips to LINK: HF AGENT.