Architecture

Single Cloudflare Pages project. Static dashboard + a Pages Function (Workers runtime) for the API. No external state.

Browser
   │
   │  GET  /                  → static index.html (this page)
   │  POST /api/research-lab  → Pages Function → OpenRouter
   ▼
Cloudflare Pages: research-lab
   ├── index.html
   ├── _headers
   └── functions/api/research-lab.ts

1. Type query — token counter (chars/4) live; buttons unlock at ≥5 chars + model catalogue loaded.
2. Click ⚡ Analyse — runs rule-based preflight in the browser. Zero network. Result populates Intent · Complexity · Tokens · Optimised Prompt · per-model cost chips.
3. Edit optimised prompt (optional) — overrides the raw query on Run.
4. Click ▶ Run — POST to /api/research-lab with selected models + memory context.
5. Pages Function fans out — Promise.allSettled hits OpenRouter once per model in parallel. Failures isolated.
6. Per-model parsing — regex captures Step N: lines; split on ## Final Answer; cost computed from local pricing table × provider's token counts.
7. Response renders — model cards with collapsible Steps + Chain of Thought, final answer, token counts, per-card cost.
8. Session pushed to localStorage['rl_h'] (FIFO, cap 50). Click in History to replay.

latencyEdge round-trip ~10 ms. Run dominated by slowest model in the fan-out (typically 5–30 s).

The ⚡ Analyse path is 100% browser-side. No LLM call, no cost, instant.

Intent — first matching keyword group wins, in this order. All matches are whole-word, case-insensitive (regex word boundaries) so derivation doesn't trigger derive.

Complexity score (1–10) — additive heuristics:

• Base: +2
• Word count >30 / >60 / >120 → +1 each tier (max +3)
• Multiple question marks → +1
• Contains acronym (3+ caps) → +1
• Contains numeric with unit (e.g. 15kg, 5%) → +1
• Contains code block, =>, function, or class → +2
• Intent in {mathematical, technical_deep_dive, experimental_design} → +1
• Intent is factual_lookup → -2 (with floor of 1)

Score maps to: ≤3 simple4–6 medium≥7 hard

Model suggestions by complexity:

• simple → gpt-oss-120b:free
• medium → gpt-oss-120b:free, gemma-4-31b-it:free
• hard → gemma-4-31b-it:free, gpt-oss-120b:free, nemotron-nano-12b-v2-vl:free

All three models are on OpenRouter's :free tier — $0 per query. Subject to free-tier rate limits.

Attached images are routed to vision-capable models as image_url content blocks; text-only models receive the prompt unchanged.

GET /api/research-lab — health check, returns service metadata.

POST /api/research-lab with one of two actions:

// list available models + pricing
{ "action": "list_models" }

// fan out a query to N models in parallel
{
  "action": "run",
  "query": "...",
  "models": ["openai/gpt-oss-120b:free", "nvidia/nemotron-nano-12b-v2-vl:free"],
  "sessionContext": "optional memory blob",
  "image": "data:image/jpeg;base64,..."   // optional; only sent to vision-capable models
}

Response shape for run:

{
  "outputs": [
    {
      "modelId": "...",
      "modelName": "...",
      "provider": "...",
      "answer": "...",
      "reasoning": "...|null",
      "steps": ["Step 1: ...", ...],
      "sawImage": true,
      "inputTokens": 0,
      "outputTokens": 0,
      "cost": 0,
      "error": false
    },
    ...
  ],
  "totalCost": 0,
  "query": "..."
}

Persistence is per-browser-per-device. Clearing site data wipes history + memory. There is currently no cross-device sync.

Infra + per-query cost: $0/mo. Free-tier rate limits apply at OpenRouter (currently ~20 req/min and a daily cap; check the OpenRouter dashboard for current values).

• Token counter is char/4 — approximate. Provider's usage.prompt_tokens is authoritative and shown post-Run.
• 4096 max output tokens per model call (set in API). Long answers may be truncated.
• No retry on provider error — failed models return inline error cards. Re-Run to retry that model.
• Free-tier rate limits — OpenRouter throttles :free models per-minute and per-day. Heavy use will see 429 errors that surface as inline error cards.
• Image size — dropped/pasted/uploaded images are base64'd inline in the request body. Stay under ~5 MB to avoid hitting OpenRouter request-size caps.
• Vision routing — when an image is attached, it's only sent to vision-capable models (Nemotron, Gemma). Text-only models still receive the prompt.
• History capped at 50 — oldest entries fall off automatically.

• Origin — Base44 monolithic Deno function at scientist-1-d28c6e14.base44.app/functions/researchLab.
• Migrated to Cloudflare Pages + Pages Functions, single-account, no Supabase needed.
• Removed dead import @base44/sdk + body.apiKey footgun before porting.
• LLM preflight stripped — replaced with deterministic browser rules.