Link Hunter is an AI-powered Chrome Extension designed for makers, crafters, and DIY enthusiasts. It allows users to right-click any craft, hobby, or handmade item image on the web and instantly discover the best YouTube tutorials on how to make it.
Instead of guessing keywords, Link Hunter uses a Vision-Language Model to analyze the image, figure out the exact materials and techniques required, and serves up the highest-quality, most relevant video guides directly in your browser.
To use Link Hunter on your own machine, follow these simple steps:
Click here to download the extension from Chrome Web Store
You're all set! Right-click any DIY image on the web and select "Search with LH" to see the magic happen.
All backend traces and multi-stage LLM pipelines are monitored using LangSmith.
| Metric | Measured Value | Description / Scope |
|---|---|---|
| P50 Latency | ~4.903s |
Median end-to-end trace latency across both LLM stages |
| P99 Latency | ~7.924s |
Worst-case tail latency (complex prompts / cold starts) |
| Error Rate | < 0.1% |
Production trace execution failure rate |
| Avg. Output Tokens | ~55 tokens |
Average response size per Trace |
| Pipeline Stages | 2 LLM Calls |
Stage 1: Visual data extraction → Stage 2: Search query curation |
Note: Metrics reflect production telemetry captured across live requests.
Link Hunter operates on a streamlined, AI-first pipeline, bridging a lightweight browser extension with a high-performance, asynchronous FastAPI backend to deliver hyper-relevant YouTube tutorials.
-
Unified Asynchronous Backend: The architecture has been migrated to a pure Python environment powered by FastAPI. This eliminates the overhead of polyglot child-process spawning, providing a unified, low-latency API capable of executing heavy machine learning workflows and handling concurrent web requests natively.
-
Two-Stage Vision-Language Pipeline: Image analysis is driven by Google's Gemma-4-31b-it model orchestrated via LangChain. To bypass strict anti-bot and hotlinking protections (e.g., on Pinterest), the backend dynamically fetches and Base64-encodes target images before transmitting them to the LLM.
- Stage 1 (Visual Routing & Extraction): Analyzes the image to immediately reject noise (mass-produced commercial items, landscapes) and extract critical DIY metadata (core subject, crafting technique, primary materials).
- Stage 2 (Intent-Driven Query Generation): Synthesizes the extracted metadata to formulate a highly specific, human-like search string that precisely answers the semantic intent of the user.
-
Bulletproof Structured Extraction: The AI pipeline enforces strict Pydantic schemas to guarantee predictable JSON outputs. To ensure zero downtime from LLM hallucinations or conversational "chatter," the system features a custom, multi-layered regex fallback parser that surgically extracts nested JSON arrays and objects, ensuring the pipeline never crashes during data transfer.
-
Semantic Re-Ranking Engine: To solve YouTube's issue of broad or irrelevant keyword matching, Link Hunter implements a custom semantic re-ranking step. It generates dense vector embeddings of retrieved video titles using Google’s Gemini embedding model (
models/embedding-001viaGoogleGenerativeAIEmbeddings) and computes the Cosine Similarity against the target query vector. Videos failing to meet the strict semantic threshold are pruned from the final recommendations.
- Prompt Engineering & A/B Testing: The query generation pipeline was rigorously optimized through structured A/B testing across 4 distinct system prompt variants. Evaluated on a curated subset of our golden dataset, Prompt V1 (Strict/Concise) initially led the pack, achieving an 83% Average Retrieval Accuracy alongside a flawless 100% Routing Accuracy (correctly distinguishing DIY vs. non-DIY images). Prompt V3 (Guided Natural) followed closely with 77% average retrieval accuracy and 100% routing accuracy.
- LLM-as-a-Judge Evaluation Framework: To ensure production-ready reliability, the pipeline was benchmarked against the complete golden dataset using an automated "LLM-as-a-Judge" architecture. This judge model autonomously graded retrieval quality by strictly matching the top retrieved YouTube video titles against expected rubric descriptions (Precision@5). In this comprehensive full-scale evaluation, System Prompt V3 demonstrated superior generalization and human-like query formulation, ultimately standing out with a sustained 83% Overall Accuracy.