dzen.

Dzen Chat technical specification

An implementation reference for security, platform and engineering teams: what runs inside Dzen Chat, how knowledge is processed, where data lives and which controls constrain each answer.

Retrieval-Augmented Generation: answer with retrieved evidence.

RAG stands for Retrieval-Augmented Generation. Before an AI model writes an answer, the system retrieves relevant fragments from your approved knowledge base and adds them to the model’s working context.

Find the evidence

Dzen Chat searches the indexed website and documents for the fragments that best match the question. It does not send the whole knowledge base to the model.

Build a bounded context

Selected sources, dialogue tail and answer policy are assembled into a controlled context. Ranking and relevance rules decide what is useful enough to include.

Explain, then cite

The chosen AI provider writes a response from that context. Dzen Chat preserves the source payload so the interface can link the visitor back to the underlying material.

Follow the data through the system.

The interface mirrors the separation between background indexing and the live chat request.

1SourcesSites · files · knowledge systems · APIs
2Crawler & extractorDiscovery · fetch · extraction · change refresh
3Index pipelineChunk · deduplicate · embed
4Knowledge storePostgreSQL · pgvector · metadata

Initial indexing and source updates run in background workers through Redis queues.

How an answer context is constructed.

For each question, Dzen Chat combines relevant dialogue with fresh evidence from the knowledge base.

Does the annual plan include additional team members?
Vector + full text + graph → fusion → rerank → 4 sources selected
Conversation tail + selected evidence + answer policy within a bounded context
It includes the base team allocation. Additional seats can be added from workspace settings.Pricing and plan management · source cited

Where do I add seats? · Can I change the billing owner?

From raw content to a maintained knowledge index.

Discover and fetch

Sources can be websites, sitemaps, documents, knowledge systems and API connectors. The crawler refreshes changed material asynchronously.

Extract useful structure

HTML and supported documents become usable content with headings, tables and metadata.

Reject noise and repeats

Content hashes and chunk hashes remove repeated templates and duplicate fragments.

Chunk with context

Content becomes retrievable fragments with section paths, entities and document metadata.

Embed and store

Workers generate embeddings and persist chunks, vectors and metadata in PostgreSQL with pgvector.

Observe and refresh

Source, crawl and embedding states remain visible to administrators.

Make Dzen Chat feel like a native part of your product.

The administration area gives your team control over the widget: colours, activation triggers, suggested questions, conversational tone, welcome messages and contextual invitations.

Open Dzen Chat widget with welcome message, pinned guidance and message composer
Public Dzen Chat widgetThe visitor-facing interface: configurable welcome state, contextual guidance and a conversation composer.

Start the right conversation from page context.

Configure visual language, welcome, suggested questions and activation rules to fit the page and visitor’s task.

Looking at plans? Ask Dzen Chat about seats, billing or annual terms.

Contextual invitationA trigger can appear after a visitor reaches a defined page section and open the conversation around its content.

LDAP & Active Directory for private deployments.

Administrators can authenticate through a company directory, with optional group rules, service-account lookup and TLS settings.

Search uses evidence, not just one scoring method.

Semantic retrieval

Question and knowledge fragments are represented as vectors, finding relevant material despite varied wording.

Lexical retrieval

Full-text search scores titles, URLs, headings and content for exact terminology.

Evidence selection

Candidate lists are fused and reranked before a compact context is built.

Graph retrieval: use relationships between knowledge sources.

Graph traversal surfaces supporting context across entities and documents, then combines it with vector and full-text candidates.

Choose the AI connection that fits your environment.

Keep retrieval and knowledge under Dzen Chat control while choosing a generation provider for your language, policy and infrastructure requirements.

  • OpenAI
  • GigaChat
  • YandexGPT
  • OpenRouter
  • OpenAI-compatible
  • Anthropic
  • Gemini
  • xAI
  • Mistral AI
  • NVIDIA NIM
  • Perplexity
  • Hugging Face
  • DeepSeek
  • Qwen
  • Baidu AI
  • Doubao
  • SAP AI
  • Cloudflare Workers AI
OpenAI-compatible

BYOM · Bring your own model.

Private deployments can connect to your model endpoint or enterprise AI gateway; Dzen Chat provides retrieval context, orchestration and visitor experience.

Guarded against unexpected requests and unsafe output.

Guardrails constrain the conversation before and after generation.

Grounded answers

The model receives a bounded evidence set and may cite only supplied sources.

Scope boundary

Prompt checks keep the assistant within the configured product, service and knowledge-base scope.

Unsafe request controls

PII, moderation, prompt-injection and jailbreak checks can stop unsuitable input.

Output review and auditability

Output checks apply the same boundary, while blocked events remain available for review.

A complete RAG system inside your perimeter.

Dzen Chat can run in your infrastructure: widget, administration, index workers and knowledge store remain inside a controlled network boundary.

Customer network

01

Public edge

Widget and API entry points.

02

Application

Administration and chat orchestration.

03

Workers & queues

Crawling, extraction and indexing.

04

Knowledge store

PostgreSQL, pgvector and metadata.

05

Model gateway

Approved generation provider route.

Your network controls

Deployment can follow your network, identity and logging requirements.

Scoped operations

We size the topology for your sources, traffic and availability needs.

Connect the knowledge systems your team already uses.

Start with the most useful sources, then add integrations deliberately.

Use your existing source of truth.

Dzen Chat can ingest websites, documents and knowledge-system APIs through supported connectors and scoped webhooks.

  • Content & documentation
  • Zapier
  • Make
  • n8n
  • Any webhook
  • Product & support
  • Notion
  • Google Drive
  • Amazon S3
  • SharePoint
  • Confluence
  • Dropbox
  • Box
  • OneDrive
  • Slack
  • Microsoft Teams
  • Custom systems
  • Zendesk
  • Intercom
  • Freshdesk
  • HubSpot
  • Salesforce
  • Jira
  • Linear
  • Contentful
  • Strapi
  • Sanity
  • WordPress
  • Webflow
  • Shopify
  • ReadMe
  • GitBook
  • Docusaurus
  • Custom CMS
  • GitHub
  • GitLab
  • Bitbucket

Review the implementation with your technical team.

Bring your requirements for security, deployment and knowledge sources. We will map a practical Dzen Chat starting point.

Talk to us