Vehicle specification queries answered accurately across 163 languages. QA-gated retrieval grounded in a single structured source of truth.
HDFC ERGO needed vehicle intelligence that could answer precise specification queries across their entire product catalogue - and answer them correctly, in the user's language, every time. Standard chatbot approaches failed: the same query in different languages returned contradictory results, creating a trust problem at scale.
Manual escalation was the only reliable fallback. Every spec-heavy query the chatbot got wrong became a support ticket. At HDFC ERGO's scale - one of India's largest general insurers - this was a volume bottleneck with a measurable cost.
The naive approach to multilingual RAG is post-hoc translation: generate a response in one language, then translate it. This works for general prose but fails for structured specification data, where terminology must be exact. A wheel size or engine displacement doesn't translate - it must be retrieved correctly in the target language from the start.
The system treats all 163 languages as first-class targets throughout the retrieval and generation pipeline. Language-aware chunking and per-language factual validation are built into the retrieval process itself, not bolted on afterward.
A RAG-based vehicle intelligence assistant grounded in a curated specification database with image-linked attributes. Every response is anchored to a single structured source of truth - no speculative answers on spec queries, no hallucinated specifications.
QA-gated retrieval validates lookup quality before generation. If retrieval doesn't meet quality thresholds, the generation step doesn't run. This gate is what makes the ~97% accuracy number hold at scale - errors are caught before they become delivered answers.
Dynamic data lookup operates across all 163 languages with localised responses grounded in the same structured data. The specification database acts as a single canonical source regardless of which language the query arrives in.