Documentation

INFEREN

Clear guides, reference pages, and operational notes in one place.

Ready to read

Getting Started

This guide covers:

  • Admin workflow: create and maintain knowledge bases, validate via chat, and monitor the system.
  • Developer/operator setup: environment variables, URLs/base paths, and common troubleshooting.

1) Where to Go in the UI

Admin UI

  • Knowledge bases: /admin/knowledge-bases
  • Sessions: /admin/sessions
  • Tracing: /admin/tracing
  • Dashboard: /admin/dashboard

AI Chat UI

  • Chat home: /ai-chat
  • KB-scoped chat: /ai-chat/{kb_id}

Documentation

  • Docs home: /docs

2) Typical Admin Workflow (End-to-End)

Step A — Create a Knowledge Base

Go to Knowledge Bases and create a new KB using one of the supported modes:

  • URL ingestion (common): provide a source URL, optional name/description, optional ingestion preset.
  • File upload (if enabled): upload a file and create a KB from it.

What to expect:

  • The KB is created immediately, but content ingestion happens in the background.
  • The KB status changes over time (queued → scraping → indexing → ready).

Step B — Wait Until the KB Is Ready

KBs move through statuses such as:

  • queued: waiting to start
  • scraping: fetching/extracting content
  • indexing: chunking/embedding content
  • ready: retrieval + chat are available
  • failed: a processing error occurred

Operational rule:

  • Treat ready as the only safe state for serious chat testing.

Step C — Review Documents

Open the KB detail page to:

  • confirm documents were ingested
  • preview content (sanity-check extraction)
  • delete or re-add problematic documents

Step D — Configure the KB

Typical configuration fields include:

  • display name + description
  • language (affects UI direction and sometimes processing defaults)
  • theme color + icon
  • source visibility in chat (show/hide citations)

Step E — Test in AI Chat

Test from:

  • the KB list “Chat” action, or
  • /ai-chat/{kb_id} directly

During testing:

  • ask simple “can it find the content?” questions first
  • verify sources/citations are correct (if enabled)
  • only then move to harder queries

Step F — Review Sessions and Traces

Use:

  • Sessions to review stored conversations (questions, answers, feedback, sources).
  • Tracing for deeper debugging (retrieval details, timing, scores, pipeline steps).

3) Developer / Operator Setup

Environment variables

The UI and scripts commonly rely on:

  • APP_NAME
  • API_BASE_URL
  • WEB_HOST

Important distinction:

  • API_BASE_URL is for API fetch calls.
  • Templates should use /admin/... and /static/... for navigation and assets.

URL & base path rules (avoids slow sidebar navigation)

Rule of thumb:

  • In templates, use /static/... for scripts/styles/images.
  • In templates, use /admin/... for admin links.
  • In JavaScript, use window.API_BASE_URL when building API endpoints.

If you suspect the rule is violated:

  1. Open browser DevTools → Network.
  2. Click a sidebar item.
  3. If you see navigation to a different origin/port and lots of asset downloads, the page is being built with the wrong base.

Optional: Client-side “AJAX” admin navigation

There is client-side admin navigation support in frontend/static/js/admin-navigation.js.

  • It is opt-in via the <html> attribute data-admin-ajax-nav="true".
  • When enabled, sidebar clicks can be intercepted to load only <main> content via fetch().

Tradeoffs:

  • Faster perceived navigation after first load.
  • Requires careful page reinitialization (some pages rely on DOMContentLoaded).

If you enable it, test every admin page thoroughly.

4) Common Gotchas

“Ready” but answers look wrong

Typical causes:

  • content extraction was poor (document preview shows junk or missing sections)
  • KB language is mismatched (RTL vs LTR UI confusion)
  • sources are hidden (appears “ungrounded” when citations would clarify)

Recommended checks:

  • open KB detail → preview a couple of documents
  • run 2–3 “obvious” queries and verify retrieval
  • check traces for retrieved chunks

“Chat” works but is slow

Typical causes:

  • large KB / large documents
  • cold start (first query after deploy)
  • vector store latency
  • upstream model latency

Recommended checks:

  • compare first query vs second query (warm-up effects)
  • inspect tracing timing breakdown if available

Next Pages