My First Million × Jev
Search what was said on My First Million: hybrid recall, Jev multi-question re-rank, and jump-to-caption timestamps.
Overview
My First Million × Jev scopes retrieval to @MyFirstMillionPod. Try the live demo at /demos/my-first-million/: hybrid recall plus one Gateway evaluate pass with exists, relevance, topic, and hit questions, then a pile UI with FLIP rise, debug sheet, and caption mention chips with &t= jump links.
Problem: YouTube channel search is either generic Google results or endless scrolling the uploads tab. Founders want "that episode where they talked about X" without memorizing titles.
Built for: My First Million listeners, indie hackers, and Jev builders who want a reference channel-search stack with typed gates instead of vibes-only reranking.
First indexed on Jev Directory: 2026-09-20
Creator and team
- Name
- Rishit Patel
- Organization
- Southern East Inc.
- Handle
- @imrishit98
How Jev is used
- Role in the product flow
- Query understanding, hybrid shortlist rerank, and per-video match gates
- Primitives
- ChoiceScoreNoul
- State in
- User query, channel metadata, and a shortlist of candidate videos with title, description, tags, chapters, and transcript snippets pulled from the local catalog.
- Decision out
- Intent Choice for query understanding; boolean exists gate; per-candidate relevance, topic, and hit Score/Noul blend with configurable thresholds before results render.
- Local lexicon expansion plus Jev Choice intent (guest, series, game, vibe, other)
- Hybrid recall: BM25, TF-IDF, fuzzy, and caption-proximity lanes fused with weighted RRF (~12)
- Single experimental_evaluate: exists, rel_*, topic_*, hit_* with transcript snippets in state
- Weighted blend (0.4 / 0.25 / 0.35) with exists and combined match gates
- UI pile, FLIP rise, debug sheet, Mentioned in captions with YouTube &t= links
Jev is not a chat wrapper here. Cloudflare Pages Functions call AI SDK experimental_evaluate with typesafe-ai/jev through the Vercel AI Gateway. One Choice classifies query intent before hybrid recall expands BM25 terms; re-rank still uses the original user query. A single evaluate batches exists, per-candidate relevance and topic booleans, and hit score rubric (off-topic, partial, direct hit) with transcript snippets in state. Application code blends 0.4 / 0.25 / 0.35 and applies exists and match gates. Hybrid recall (including caption-proximity lane) stays local; Jev only sees the shortlist metadata blob.
Features and stack
Features
- Hybrid recall (BM25, TF-IDF, fuzzy to RRF) in hybrid-recall.mjs
- Query understanding before BM25 expansion
- Multi-question Jev with exists and match confidence gates
- Caption mention timestamps and Jump to mention links
- Flat-playlist ingest and optional yt-dlp enrichment for captions
- Debug sheet with understand, shortlist lanes, and full JSON
Stack
- Cloudflare Pages Functions
- Vercel AI SDK + AI Gateway
- typesafe-ai/jev
- Hybrid BM25 / TF-IDF / fuzzy / caption RRF
Pricing: Open source demo. Live Jev calls bill to your AI Gateway key; mock mode skips network.
Links
FAQ
- Does this search the whole internet?
- No. Retrieval and scoring are scoped to the My First Million uploads catalog bundled at src/data/mfm-channel-catalog.json.
- Can I run it without a Gateway key?
- Yes. Set JEV_MOCK=true in Cloudflare Pages env vars (or .dev.vars for wrangler pages dev). Mock mode skips Gateway calls and uses deterministic scores.
- Where do caption jump links come from?
- When catalog videos include SRT/VTT caption text, the demo finds mention windows and builds YouTube &t= URLs in the feature card.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- Jev vs LLM classification
When to gate with System One probabilities instead of asking a chat model to label things.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.