Inspiration
Support teams already have AI that summarizes conversations and drafts replies, but companies still struggle to answer a more important question: why do customers keep opening the same tickets? Documentation and support tickets usually live in separate systems, so knowledge teams cannot easily prove which help articles are incomplete, outdated, or contradictory.
What it does
Support Gap Radar turns historical customer tickets into a regression test suite for company documentation.
- A team uploads a vendor-neutral CSV of resolved support tickets and its current documentation in Markdown, text, PDF, or Word format.
- OpenAI embeddings group semantically similar questions and retrieve the most relevant documentation passages.
- GPT-5.6 performs a structured coverage audit and labels each question as covered, partial, missing, or contradiction.
- The visual radar shows repeated knowledge gaps and the evidence behind them.
- For a safe cluster, GPT-5.6 drafts an editable, evidence-grounded Markdown patch.
- Knowledge Replay retests the same historical questions against the current documentation plus the proposed patch, showing question-level before/after projected coverage.
- If documentation and resolved support outcomes disagree, drafting is blocked until a human makes the policy decision.
Unlike a generic support summarizer, the product creates a reusable loop: detect, patch, replay, review. It can work as an intelligence layer above Zendesk, Salesforce, Front, or any support platform that exports tickets.
How we built it
- React 19 and Vite for the interactive radar and review workflow
- Netlify Functions for server-side OpenAI calls and credential protection
- text-embedding-3-small for semantic clustering and retrieval
- Deterministic k-means and cosine similarity for repeatable grouping
- GPT-5.6 Structured Outputs validated with Zod
- Browser parsing for CSV, Markdown, text, PDF, and DOCX
- Editable Markdown export, contradiction blocking, and Knowledge Replay
How Codex accelerated the work
The majority of the project was built in one Codex task. Codex helped research the competitive baseline, turn the original documentation gap detector into a regression-testing product, generate the visual direction, implement the React and serverless workflows, create synthetic datasets, diagnose API failures, and verify the result with lint, automated tests, a production build, and a live GPT-5.6 smoke test.
Important decisions made with Codex included keeping the tool vendor-neutral, grounding drafts only in supplied evidence, blocking unsafe generation when policies conflict, storing the API key only on the server, using store: false, and reporting projected documentation coverage rather than promising ticket deflection.
Challenges
The hardest parts were defining honest coverage states, separating incomplete documentation from true policy conflicts, preventing unsupported policy invention, and making AI evaluation reproducible enough for judges to inspect. We solved these with structured outputs, evidence citations, deterministic grouping, conservative guardrails, and human review before export.
Accomplishments
- A deployed demo that judges can test without credentials
- Three synthetic sample packs
- Evidence-level before/after replay results
- Contradiction guardrails that block unsafe drafting
- Fourteen focused automated tests and a live GPT-5.6 integration smoke test
- A responsive, complete workflow instead of a static proof of concept
What we learned
Embeddings and RAG become much more valuable when they support a measurable product loop. The most important lesson was that AI should not only generate content—it should show evidence, expose uncertainty, and help a human decide whether a change is safe.
What’s next
Next steps include direct Zendesk, Salesforce, and Front imports; scheduled documentation regression runs; team review workflows; and production evaluation datasets that compare projected coverage with real ticket deflection over time.
Log in or sign up for Devpost to join the conversation.