🛠️ Tech Stack

🚀 Features

• Text → Emoji: embed a natural language query and rank ~1,800 emoji by cosine similarity in <5ms using brute-force NumPy — no vector database needed at this scale

• Emoji → Text: two-stage pipeline — deterministic metadata lookup (handles ZWJ sequences, flag pairs, skin-tone modifiers) followed by GPT-3.5 Turbo interpretation

• Two-pass diversity filter that prevents redundant results (e.g., five identical smiley faces) by enforcing one emoji per Unicode category, with automatic relaxation when categories are sparse

• SQLite embedding cache eliminates redundant API calls across sessions — zero external dependencies beyond the stdlib

• Interactive REPL with auto-detection of input type (text vs. emoji) for rapid bidirectional exploration

• Four-stage data pipeline (fetch → merge → describe → embed) that combines Unicode CLDR and Emojilib into richly described, template-augmented embeddings for ~$0.02 total build cost

🧗‍♂️ Challenges

• Discovered that embedding raw keyword lists scored significantly worse than embedding template-augmented descriptions — context engineering ('pizza — commonly used to express: cheese, food, hungry') outperformed model selection for retrieval quality

• OpenAI SDK v1.0 uses httpx connection pools that hang short-lived CLI processes; diagnosed the thread lifecycle issue and replaced single-embedding calls with direct urllib requests while keeping the SDK for batch operations

• Designed a two-pass diversity filter to handle conflicting requirements — strict category deduplication for common queries vs. relaxed backfilling for deep-category queries like 'sushi dinner' where most results share Food & Drink

• Built a 50-case evaluation suite across 10 semantic categories revealing that concrete nouns average 2.7+/3.0 while abstract emotions and sarcasm average ~2.0 — quantifying exactly where embedding-based search hits its limits

🧠 Learnings

• What you embed matters more than which model you use — template-based description augmentation improved retrieval quality more than any other single change, a lesson that carried directly into later RAG projects

• Right-sizing infrastructure is an engineering skill: 1,800 vectors × 1,536 dimensions fits in 14MB of RAM and searches in <5ms with brute-force NumPy — a vector database would add operational complexity for zero performance gain

• Cross-modal semantic search is just regular search in the right embedding space — building this at small scale made the CLIP / multimodal retrieval principle concrete before needing it in larger systems

• Even for a fun side project, a scored evaluation suite caught failure modes (sarcasm, abstract emotions) that manual spot-checking completely missed — this discipline became the foundation for later evaluation work