Polylog
Multi-agent chatbot platform with RAG, flow builder and widget
A multi-tenant, multi-agent chatbot platform. Clients sign up, upload knowledge, design conversation flows visually, embed a widget on their site, and route captured leads to Telegram, WhatsApp, email, or a webhook — all from one dashboard.
- Sector
- AI · SaaS
- Stack
- Next.js 15 · FastAPI · Elasticsearch · Stripe


- Languages (cross-lingual RAG)
- 100+
- Dockerized services
- 9
- Tenant + multi-agent
- Multi
- Grounded, cited answers
- RAG
One widget, many agents, every language
Polylog lets a business stand up a grounded, on-brand chatbot in minutes: upload PDFs, crawl a site, or paste text, design the conversation visually, and drop one script tag on the page. Every row and query is scoped by tenant, and a single widget can host several agents — say a travel consultant and a sales consultant side by side.
Answers are retrieved with hybrid search (BM25 + dense vectors + RRF) over Elasticsearch, reranked by a cross-encoder, with bge-m3 embeddings that work across 100+ languages. A grounding verifier checks every answer against the retrieved sources and blocks ungrounded claims before the visitor ever sees them.
Ship a SaaS where non-technical clients build a trustworthy chatbot themselves — accurate, multilingual, and safe from hallucinations — while keeping every tenant's data strictly isolated and the whole thing self-hostable.
A FastAPI backend with tenant-scoped Postgres + Elasticsearch, a Next.js dashboard with an @xyflow visual flow builder, a Shadow-DOM widget that streams over SSE, Stripe billing with server-side limits, and a grounding/citation trust layer. Nine services orchestrated with Docker Compose.
Everything you need in one platform
- 01
Hybrid RAG knowledge
Upload files, crawl URLs, or paste text. BM25 + dense vectors + RRF on Elasticsearch, cross-encoder rerank, bge-m3 cross-lingual embeddings.
- 02
Visual flow builder
Message, choice, capture, condition, API-call, AI-answer, handoff and intent nodes on an @xyflow canvas, with a live test-run.
- 03
Embeddable widget
Shadow-DOM isolated, streams answers over SSE, renders choice buttons, lead forms, citations, and product cards — localized to the visitor.
- 04
Multi-tenant + multi-agent
Every row and query scoped by tenant_id; knowledge and flows scoped by agent_id. One widget can host several agents.
- 05
Channels & leads
Route captured leads to Telegram, WhatsApp, email, or webhook — credentials encrypted at rest, never in env.
- 06
Grounding trust layer
A verifier checks each answer against retrieved chunks and blocks ungrounded claims, with inline citations and a guardrail audit log.
See It in Action


Built on a sharp, modern stack
A typed Python backend, a React 19 dashboard, and Elasticsearch-backed retrieval — orchestrated as nine Dockerized services with Stripe billing.
- Next.js 15
- Dashboard + landing (next-intl)
- FastAPI
- Tenant-scoped Python backend
- Elasticsearch
- Hybrid retrieval (BM25 + dense + RRF)
- Stripe
- Checkout + metered plan limits
Want a grounded chatbot for your product?
Multilingual, multi-agent chat with RAG knowledge, lead capture, and a trust layer that blocks hallucinations — embeddable on any site.
CallLens
Sales calls transcribed, scored by an LLM, searchable by meaning

Let’s talk.
Tell us what you’re building. You’ll get a reply and a first estimate within 24 hours.
Fill in the brief →