Manage / Conversations

Conversations & visitor logs

Inspect real visitor interactions with your AI assistants. Every visitor session captures full message transcripts, source citations, token consumption, and per-turn response latency.

Session architecture

How visitor sessions are generated, persisted, and tracked across widget interactions.

When an end user opens the widget on your site, the widget client establishes a unique session identifier with the WebChat AI backend. This session groups consecutive question-answer turns into a cohesive conversation thread, allowing the assistant to maintain conversational context while isolating data per visitor.

Conversation Turn Lifecycle

The 4 stages from visitor input to persisted telemetry.

STAGE 1

Visitor Message

User types question in widget; client sends payload with active session token.

STAGE 2

RAG Retrieval

Context retrieved from vector + lexical indexes; source citations assigned.

STAGE 3

Generation & Telemetry

LLM generates reply; backend records input tokens, output tokens, and latency.

STAGE 4

Dashboard Sync

Session transcripts and metrics immediately surface in the dashboard inbox.

Navigating the inbox

Searching, filtering, and reviewing visitor conversation history.

The Conversations dashboard (/conversations) provides an inbox view of all visitor interactions across your registered assistants:

Assistant Filter: Switch between all assistants or isolate conversations for a specific website or document-based assistant.
Real-Time Search: Debounced (300ms) full-text search across message content, visitor labels, and session IDs.
Status Badges: Identifies active sessions, completed threads, and partial responses where generation encountered upstream timeouts.
Pagination: Paginated at 20 conversations per page with immediate total counts and server-side navigation.

Message details & telemetry

Auditing source grounding, citations, and model performance on every turn.

Clicking any conversation opens the detailed transcript viewer. Each assistant message exposes the underlying facts and operational telemetry:

Message turn telemetry fields
Telemetry Metric	Example	Description
Source Citations	[1] /docs/pricing	Direct hyperlinks to crawled pages or uploaded documents grounding the answer.
Input Tokens	412 tokens	Prompt tokens including system prompt, grounded chunks, and conversation history.
Output Tokens	86 tokens	Completion tokens generated by the model for this specific reply.
Response Time	340ms / 1.4s	End-to-end turn duration from question submission to stream completion.
Message Status	completed / failed	Flags partial answers if generation was aborted or hit provider rate limits.

Privacy & conversation deletion

Tenant data ownership and compliance controls.

Conversations belong entirely to your tenant account. You can permanently delete any conversation session directly from the detail view:

Hard Deletion: Deleting a session immediately removes all associated user messages, assistant responses, source links, and token logs from the database.
Multi-Tenant Isolation: Conversation queries are strictly scoped by tenant_id. Cross-tenant reads or mutations return 404 or 403.
No End-User PII Required: By default, visitors are tracked with anonymous pseudorandom session tokens without requiring cookies or personal identification.

Auditing AI Quality

When visitors ask questions that trigger the ungrounded answer fallback (e.g. “I do not have enough information”), use those questions to identify missing content and upload supplemental PDFs or trigger website re-crawls.