A Chinese manufacturer, Latin American contractors, and a catalog no one could navigate
PJCS makes industrial waterproofing materials in China. Their customers are professional contractors across Mexico, Colombia, Peru, and other Latin American markets — many of whom speak only Spanish and have highly specific technical requirements: the right product for a specific substrate, application environment, and budget.
With 84 products across categories ranging from roof coatings to foundation membranes, finding the right product meant calling a rep, digging through a PDF catalog, or guessing. The language barrier between Mandarin-speaking product managers and Spanish-speaking contractors made this even harder.
They needed a system that could answer "Which product works for an exposed rooftop in a high-UV environment with a $500 budget?" — in Spanish, English, or Chinese — without requiring a human in the loop.
Hybrid search + LLM synthesis, built for three languages
The core insight: pure vector search misses exact specification matches (e.g., a specific product code or chemical name), while pure keyword search misses semantic intent ("something waterproof for basements"). We combined both with Reciprocal Rank Fusion to get the benefits of each.
Query received
User submits natural-language question in English, Chinese, or Spanish.
Hybrid retrieval
Query runs in parallel through ChromaDB (vector embeddings) and BM25 (keyword matching).
Reciprocal Rank Fusion
RRF merges and re-ranks both result sets for higher-precision candidate selection.
GPT-4o-mini synthesis
Top candidates passed to the language model with product specs as grounding context.
Response delivered
Answer returned in the user's language with product names, specs, and application guidance.
Beyond the chat interface, we built a 5-step guided product finder wizard: users select their application type, substrate, scale, and budget, and the system narrows to a curated recommendation. Two audience modes — buyer and professional installer — surface different levels of technical detail.
The system degrades gracefully: if the AI layer is unavailable, the hybrid search results are shown directly. Nothing breaks silently.
Deployed on Railway, serving three markets
Trilingual coverage
Full support across English, Chinese (Simplified), and Spanish — UI, AI responses, and product data.
Hybrid RAG pipeline
ChromaDB vector search + BM25 keyword matching merged via Reciprocal Rank Fusion for higher-precision results.
5-step product wizard
Guided discovery from application type to final recommendation, with audience mode switching.
Graceful degradation
Full functionality even when the AI layer is unavailable. No silent failures, no broken UX.
Auto-deploy pipeline
Deployed on Railway with continuous deployment from GitHub. Zero-downtime updates.
84 products indexed
Complete product catalog embedded, chunked, and searchable with full spec data and application guidance.
Next Project
DocProc PDF Toolkit