Challenge
Chat theatre is liability in customs
What problem existed?
Importers and exporters ask hard questions: classification framing, documentation checklists, AEO concepts, process education. Generic support bots invent prices, duties, and ETAs — deadly for a CHA brand. Intercom-style SaaS cannot hold Indian nuance, Neo’s own service catalog, or DPDP control without becoming a liability.
- ▸Off-the-shelf bots hallinate duties and timelines
- ▸Keys and PII in the wrong place if you ‘just embed ChatGPT’
- ▸Knowledge drifts as Neo content and public rules evolve
- ▸No disciplined path from chat interest → human sales
What we built
A native Neo brain with teeth
What PragmaLabs created
We built Neo Assist as a product package: Express API, editable Markdown knowledge base, hybrid retrieval that works without a heavyweight vector ops platform for current KB size, Gemini first with swap to OpenAI/Anthropic, prompt + output checks that refuse binding duty advice, and consent before PII storage.
Grounded retrieval
Answers cite curated Neo and educational sources — not free-ranging web fantasy.
Refusal is a feature
The assistant will not invent duties, prices, or promised timelines. Trust over glibness.
Lead + privacy
Serious enquiries become records with notify; consent and deletion paths respect DPDP pressure.
How it works
Package-first delivery
Simple workflow — how delivery was sequenced
- 1
KB & product analysis
Compared Intercom/Drift/fine-tune-only options; chose hybrid RAG + Gemini Flash for cost, control, and accuracy.
- 2
Server + retrieval
Chat, health, consent, deletion endpoints; ingest pipeline for Markdown KB.
- 3
Widget
Accessible Neo-branded floating chat, mobile ready, embed assets for site integration.
- 4
Ops controls
Provider env switches, lead notify to Neo inbox, retention-oriented design notes.
In detail
Key parts of the product
Hybrid RAG
Local lexical hybrid by default; optional Voyage embeddings path without app rewrite.
Provider interface
gemini | openai | anthropic via AI_PROVIDER — model economics without cornering Neo IT.
Guardrails
Prompt and output checks: no invented duties, binding legal advice, or fake ETAs.
Lead detection
Serious / huge enquiry signals → SQLite leads + email; Slack/Sheets hooks ready.
DPDP posture
Explicit consent before PII storage; deletion request surface; retention-minded jobs.
First-party embed
neo-assist.js / CSS assets so Neo owns the customer experience.
Technology
Technical stack
Server-side keys only. Widget is presentation; Express is the brain boundary. KB is editable Markdown so Neo ops and marketing can update facts without a deploy ceremony for every sentence. Worldwide coverage is educational by design; binding foreign counsel still escalates to humans — that is product integrity, not a bug.
- Express
- Gemini / OpenAI / Anthropic
- Hybrid RAG
- SQLite leads
- First-party widget
- DPDP consent flows
Result
What improved
- Neo can answer service and education questions without inventing duties.
- High-intent chat becomes a lead event, not a lost thread.
- Embed path exists for Neo’s digital surface as a native brand asset.
- Model vendor choice remains an env flag, not a rewrite.
Neo Assist is the proof that vertical AI must be more stubborn than it is chatty. In customs, the answer ‘I will not invent that number’ is worth more than a fluent wrong one.
“A CHA bot that invents duty rates is not innovative — it is a compliance incident waiting for a screenshot.”
Want something like Neo Assist for your business?
Book a free discovery call or a deeper working session — we start from your business problem, not a template.