Meaning, not keywords
The same embedding model converts catalogue records and incoming comments into comparable vectors.
PRODUCT + IMPLEMENTATION
How one comment
becomes the best options.
The live product compares meaning, social dynamics, and tone—not just words. It retrieves 30 likely matches from 1,000 memes, ranks the best three, then explains each choice.
Try a commentLive now: 1,000 reactionsSemantic retrieval · three scored options · serious-content guard · feedback
PERSONAL TOOL
THE LIVE 1,000
Retrieve. Classify. Explain.
Each reference has a name, meaning, relational pattern, example, tags, provenance, and media status. Vector search narrows 1,000 records to 30 before a dedicated classifier ranks the best three.
- 01Validate
Trim and bound the comment. Treat it as untrusted text.
- 02Retrieve 30
Use BGE embeddings and Vectorize to find the closest semantic matches.
- 03Guard and rank
Jev checks sensitive help-seeking text and ranks the three closest social reactions.
- 04Explain three
Llama writes specific explanations and fit scores without changing the ranked IDs.
THE RETRIEVAL SYSTEM
The path scales to 30,000.
At 3,000, retrieval keeps the prompt bounded. At 30,000, a first routing decision chooses meme, movie dialogue, or no reaction before searching the matching corpus.
COMMENTWhat happened
→
EMBEDMeaning vector
→
RETRIEVETop 30
→
RERANKBest 3
A bounded shortlist
The vector index returns 30 likely candidates from hundreds or thousands of records.
Tone and relationship
A fast classifier checks whether humour belongs and ranks candidates by social fit rather than keyword overlap.
Evidence, not vibes
Yes/no feedback is retained for evaluation; it does not silently retrain the live model.
THE EXPANSION PLAN
30 → 300 → 1,000 → 3,000 → 30,000
300 references are live in the personal tool. Evaluation now guides metadata cleanup and the next expansion.
1,000 / 3,000 live“I love you 3,000” is the north star.
2,700 memes remaining · 10% complete
- 30Seed controlLIVE
Curated metadata, direct all-candidate comparison, live public feedback.
- 300Common reactionsLIVE
Semantic retrieval, scored multi-option results, searchable collection, and feedback.
- 1KBroader vocabularyLIVE
Seven hundred new records add long-tail reactions with searchable previews and meaning-specific metadata.
- 3KI love you 3,000BUILDING
The next 2,000 will expand the long tail while the same retrieval and safety gates protect relevance.
- 30KMovie dialogue reactionsNEXT CORPUS
Route to a separate dialogue corpus first, then retrieve and rank within it so memes do not crowd out the right line.
THE EVAL DATASET
110 held-out situations guide the next jump.
The original and expansion sets cover 50 situations. The 1,000-meme set adds 60 cases: 45 humour and 15 no-meme prompts, with labels independently audited twice and still pending owner review.
- Top 1Did the first choice land?
- Top 3Was any shown option sendable?
- AbstentionDid it stay serious when humour did not belong?
- CalibrationDid low confidence actually mean weaker results?
Latest draft: the 1,000-meme path reached 96% retrieval, 87% top-one relevance, 96% top-three relevance, and 93% correct serious-content abstention. Labels remain pending owner review.
THE EXPANSION RULE
Bigger is useful only if the results stay good.
The personal catalogue stays available while evaluation tracks top-three relevance, correct no-meme decisions, valid IDs, and latency. Comments are processed by Cloudflare Workers AI and classifier.dev; previews load from their listed source hosts.