TL;DR
- Off-the-shelf (Intercom Fin, Drift AI, ChatBase): live in days, ₹15-50K/mo, limited customisation.
- Custom RAG: 2-4 weeks build, ₹1.5-4L upfront + ₹15-40K/mo running, fully customisable.
- Pick custom when: India-specific (Hindi/regional), WhatsApp-deployed, integrated with internal CRM/Tally, multi-tenant.
- Pick off-the-shelf when: standard support FAQs, English-only, quick proof-of-concept.
What "RAG chatbot" actually means
RAG = Retrieval-Augmented Generation. Your chatbot doesn't "know" your content — it retrieves relevant chunks at query time and feeds them to an LLM that generates the answer. Means it can answer about your products/policies/docs, not just generic knowledge.
Three components: (1) ingestion + embedding pipeline (turns docs into searchable vectors), (2) retrieval (find top-k relevant chunks for query), (3) generation (LLM writes the answer using retrieved context).
Off-the-shelf options in 2026
| Tool | ₹/mo | Strengths | Limits |
|---|---|---|---|
| Intercom Fin | ₹50K+ | Polished UX, deep Intercom integration | Expensive, English-first |
| ChatBase | ₹3-15K | Cheap, fast setup | Limited customisation |
| Drift AI | ₹40K+ | Sales-focused, lead-qual built in | B2B SaaS focus, not SME |
| Cohere / OpenAI Assistants | Per-token | Best models, dev-friendly | You build UX |
Custom RAG — what it costs and what you get
Build cost: ₹1.5-4L for 2-4 week build (depends on doc volume + integrations).
Running: ₹15-40K/mo (LLM inference + vector DB hosting + admin UI).
What you get:
- Trained on YOUR content (PDFs, website, knowledge base, past tickets)
- Hindi / regional language support
- Deployed on website AND WhatsApp Business API
- Integrated with your CRM (escalates to human, logs against contact)
- Fine-tuned tone (formal/casual, brand voice)
- Source citations ("based on our policy doc, page 3")
- Continuous learning loop (you mark answers as good/bad)
The decision framework
- How standard is your support? If 80% of queries are FAQs from a known doc set → off-the-shelf works. Long-tail edge cases → custom wins.
- What's your monthly conversation volume? <500/mo → off-the-shelf cheaper. >5K/mo → custom pencils out.
- WhatsApp-required? Off-the-shelf rarely deploys to WhatsApp Business API directly. Custom = yes.
- Multilingual? Custom with Qwen / fine-tune. Off-the-shelf rarely good in Indian languages.
- Internal integrations? If chatbot needs to fetch order status from your system, lookup customer in CRM, etc. — custom required.
For SME clients with >1K monthly conversations, custom RAG is usually the right call. We deploy on your AWS/DO with Llama 3.3 + open-source vector DB (Qdrant/Pinecone). Chat widget on website + WhatsApp BSP integration + admin dashboard. Total: 2-3 weeks. See AI builds →
FAQ
Will it hallucinate?
Less than pure LLM, more than zero. Mitigations: strict prompts ("only answer from context"), source citation in response, confidence thresholding (escalate to human below threshold).
How much content do I need to train it?
RAG works with as little as 50 pages. Quality scales with quantity up to ~5,000 pages, then plateaus. Focus on quality (curated, current, accurate) over quantity.
Last reviewed: 7 April 2026.
Want this built for you?
Talk to Kashvi — 30-min call, honest assessment, no pitch deck.