[im-rk]

What this is

An end-to-end audit of the live site (Oct 3), the fixes it called for, and one new feature built on the same two engines (Cloudinary transformations + one citation-grounded AI call). Spec and evidence: docs/phases/PHASE-18-query-understanding-learning-paths.md.

Audit findings → fixes

Finding (real, reproduced)	Fix
"What is the bias-variance trade-off?" returned not covered against a lecture titled The Bias–Variance Trade-off. Keyword search never saw the lecturer's words	Question understanding: when keyword retrieval is thin (or the question is in another script), one AI call rewrites it into the library's own vocabulary; keyword hits keep priority. The prompt now says honestly when a term isn't used and explains the related ideas with citations
Hindi questions silently dropped when a model was rate-limited	Same step; the answer is now written in the asker's language (the system prompt names it)
A dead model (Gemini 429 quota, seen live) cost every request its full timeout	Circuit breaker: quota/auth/overload failures skip that model for 20–60 s
A slow provider could push a request past maxDuration	Whole-chain deadline shared by every attempt
"Is AI actually working?" had no answer	GET /api/health?deep=1 makes one tiny real call per model
"regularization" found no moment, hiding the session about it	Find now also shows Sessions about this (title, speaker, chapters)
Hindi subtitles existed for 1 of 6 sessions	Backfilled the other 5 with Cloudinary explicit (data fix, no code)

New: Learning Paths (/learn, /p/[id])

A topic becomes an ordered 3–5 step micro-course across the library, edited into one Cloudinary reel. Every step is a real retrieved moment with a "why this comes next". validatePath() drops anything the model wasn't given, and returns no course rather than padding (same discipline as Ask). Stored and shareable with a Cloudinary share card. "Go deeper" under every answer opens one.

Verified on real data

• Bias-variance question → answered honestly with 6 citations (was not_found); off-topic questions still not_found
• Hindi → Hindi answers citing English lectures; 5-clip reel
• "Optimizers for training neural networks" → 4 steps from 2 lecturers, 65 s reel that renders (HTTP 200)
• A real Gemini 429 during testing: the breaker skipped the dead model and Ask kept answering

Checklist

• pnpm lint, pnpm typecheck, pnpm test (108 passing; new: breaker, query plan, path validator, answer language)
• migrations/006_learning_paths.sql (already applied to the shared Neon database)
• Docs: phase 18, API, implementation plan
• No secrets in this diff

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
pravaha	Ready	Preview	Oct 3, 2026 4:11am UTC