ClearTerms-AI

A focused, production-ready way to turn dense Terms & Conditions into clear, actionable understanding. Users can import PDFs/DOCX/plain text/URLs, get a tight bullet summary of what matters, see risk indicators, and ask grounded questions — all powered by serverless AI that runs on demand with zero long-running infrastructure.

---

Live: https://clearterms-ai.vercel.app
Deployed on Vercel with on-demand serverless AI.

---

What it delivers

• Instant, plain-English summary
  Compact bullets that surface the essentials: permissions and restrictions, pricing and fees, disclaimers, termination, governing law, privacy/data handling, IP, and AI/model-training language.

• Ask anything (grounded)
  User questions answered strictly from the provided document; if the fact isn’t present, the response states that it’s not in the text.

• Risk indicators
  Deterministic tags (low/medium/high) for patterns like arbitration, class-action waiver, indemnification, limitation of liability, auto-renewal, price changes, broad licenses, data sharing/retention, biometrics, and more.

• Reading efficiency
  Word count, estimated read time, and “time saved” versus reading the entire document.

• Flexible import
  PDF (pdf.js worker), DOCX (mammoth), raw text, or URL (HTML → text), with robust clean-up to reduce noise before modeling.

---

How it works (expanded)

1. Import

   • Accepts PDF/DOCX/text/URL.
   • URLs: strip HTML/scripts/styles and keep visible text.
   • Files: extract raw text (pdf.js/mammoth).

2. Normalize & pre-clean

   • Unicode normalization; de-hyphenation across line breaks.
   • Remove list/heading tokens at line starts (e.g., 3.1, (A), iv.).
   • Drop cross-references and admin noise (URLs, VAT, addresses).
   • Stitch sentence continuations into full sentences.
   • Enforce a size cap for responsive requests.

3. Chunking for AI

   • Split into ~8k-character chunks with 300-character overlap.

4. Serverless AI: Summary

   • The frontend calls a serverless function that invokes Gemini 2.5 Flash with a strict, JSON-only prompt.
   • If the model output is malformed/empty, deterministic fallbacks convert sentences into bullets so users still get a result.
   • UI capitalizes the first character on display/export for polish (source data remains untouched).

5. Risk pass (deterministic)

   • Regex rules run against bullets to attach risk labels and a top severity.
   • No LLM involved — fast and predictable.

6. Serverless AI: Grounded Q&A

   • Simple retrieval ranks relevant chunks and builds a focused context window.
   • Gemini is invoked in a serverless function to answer only when supported by that context; otherwise it clearly notes the information isn’t in the document.
   • Responses are concise and tied to the provided text.

7. Client-side UX & persistence

   • Summaries are cached locally for a smoother experience.
   • Export to TXT/Markdown with display-level capitalization.

---

Serverless AI calls

• Each AI action is an isolated function call

  • Analyze → function invokes Gemini and returns JSON.
  • Ask → function performs lightweight retrieval + Gemini and returns a compact answer.

• Operational benefits

  • Scales automatically with concurrent users.
  • Usage-based execution (no idle processes).
  • Secrets (API keys) remain server-side in environment variables.
  • Same-origin /api/... calls — no CORS overhead.

---

Tech highlights (expanded)

• UI/UX

  • React + Vite with a clean, responsive interface.
  • Display-level capitalization for consistent style without mutating data.
  • Reading stats and time-saved metrics.

• Text import

  • pdf.js worker via Vite worker URL (stable in dev/prod).
  • mammoth for robust DOCX → text.
  • URL fetch with HTML stripping and safe fallbacks.

• Pre-processing

  • Targeted clean-up: token/heading removal, cross-ref stripping, sentence stitching, whitespace tidy.
  • Size limits keep LLM calls snappy.

• LLM usage

  • Gemini 2.5 Flash for summary & Q&A (low latency, temperature 0).
  • Strict prompts; JSON for summaries.
  • Post-processing and layered fallbacks to avoid empty/invalid outputs.

• Retrieval for Q&A

  • Fast, dependency-light scoring that blends approximate semantics with lexical overlap.
  • No external databases; computed on the fly before each call.

• Risk tagging

  • Pure regex rules with curated patterns and category notes.
  • Consistent low/med/high signals without model variability.

• Security & privacy

  • API key lives only in serverless environment variables; never exposed to the client.
  • No server-side persistence of user content.
  • Clear in-app notice that results are assistive.

---

Deployment

Hosted on Vercel, with AI calls executed via serverless functions at /api/....

• The AI key is stored as a deployment environment variable and isn’t exposed in the browser.
• Deployments can be triggered from the Vercel CLI or dashboard; a primary domain serves the current production build.
• Client-side routing supports direct reloads on routes like /ask and /summary.

---

Local development (quick)

npm i
npm i -g vercel
vercel login
vercel link
vercel env add GOOGLE_API_KEY development
npm run vercel:dev
# open http://localhost:3000 and http://localhost:3000/api/health

In production, a SPA catch-all rewrite enables deep links like /ask or /summary on reload. In local dev, a dev-specific config omits that rewrite so Vite’s assets load correctly.

---

User flow

1. Upload / Paste / URL → Analyze
2. Review the AI Summary and risk labels
3. Ask follow-up questions to drill into specifics
4. Export the summary (TXT/Markdown) if needed

---

Note

ClearTerms-AI is designed to help you understand documents; it isn’t a substitute for legal advice.