A B2B lead chatbot in 7 languages, live in 4 countries — built in a weekend

Most B2B companies handle product inquiries by email, with a response time measured in hours. For a manufacturer selling across four European markets in six languages, that gap between question and answer is where leads are lost. Optimus was built to close that gap.

The architecture: RAG over real product documents

Optimus (V6) runs on Dify, with MiniMax-Text-01 as the LLM. Its knowledge base consists of 33 indexed documents — product sheets, technical specifications, installation guides, FAQ and subsidy information per country. Documents are chunked and embedded using bge-m3 (NVIDIA NIM), stored in pgvector, and retrieved at query time using semantic search. The chatbot never hallucinates a spec it wasn't given: if the answer isn't in the indexed documents, it says so and offers to connect the visitor with the sales team.

7 languages, 4 domains, one deployment

The same Dify instance serves all four country domains. Language detection is automatic — the chatbot responds in the visitor's language without configuration. All seven language variants (French, English, Dutch, German, Italian, Luxembourgish and a formal French variant for Switzerland) were validated and tuned in a single weekend.

Lead qualification and automated follow-up

When a conversation reaches a qualification threshold — product interest confirmed, contact details provided — Optimus triggers an automated email sequence. Two templates per domain, adapted to the local market context (Belgian subsidies, Luxembourg primes, Swiss regulations). The sales team receives a structured lead summary; the visitor receives a follow-up within seconds, not hours.

The numbers that matter

Average response latency: 1.3–1.7 seconds. Availability: 24/7 across all four domains. Languages handled: 7. Documents indexed: 33. Lead emails triggered automatically: yes, from minute one. Time a sales team member spends on repetitive product questions: trending toward zero.

What a weekend build actually means

The speed was possible because the hard parts — RAG pipeline, embedding infrastructure, multilingual prompt design, email automation — were already built and tested on previous projects. A weekend build is the dividend of six months of compounding infrastructure. The client got a production-grade chatbot, not a prototype; it has operated continuously since deployment with no manual intervention.

Token cost vs. Belgian human equivalent

A Belgian digital agency building an equivalent system — multilingual chatbot, RAG over 33 documents, automated lead email sequences, 7 language variants, 4-country deployment — would typically quote €20,000–35,000 for design, development and QA, with an 8–12 week delivery timeline. Ongoing: a part-time customer support role to handle after-hours inquiries costs €1,400–2,000/month gross in Belgium, plus the implicit cost of delayed lead response during nights and weekends. The AI infrastructure: MiniMax-Text-01 inference for approximately 500 conversations/month at €0.07/M tokens ≈ €0.04/month (essentially free at this volume); NVIDIA NIM bge-m3 embeddings for document retrieval ≈ €3–8/month; Dify self-hosted, zero licence cost. Total monthly AI cost: under €15. The weekend build timeline (vs. 8–12 agency weeks) saved an estimated €25,000 in development cost and several months of time-to-market. The chatbot's 24/7 availability replaces after-hours support estimated at €1,500/month — equivalent to €18,000/year in avoided labour against under €180/year in token costs. Cost ratio: roughly 100:1.