QuackbackQuackback
Navigation

Use AI-powered features

Let AI handle the repetitive work. Quackback uses large language models and vector embeddings to summarize feedback, detect duplicates, and triage incoming posts so your team can focus on decisions.

Overview

Quackback includes four AI capabilities:

FeatureWhat it does
Post summariesGenerates a short summary of user comments on a post
Duplicate detectionFinds existing posts that match new feedback using vector similarity, full-text search, and LLM verification
AI suggestionsTurns integration feedback (Slack, API) into actionable suggestions: merge, vote, or create
EmbeddingsVector representations of post content, used for similarity search

All features require an OpenAI-compatible API key. Without one, AI features are inactive and the UI elements don't appear.

Configure AI

Set the OPENAI_API_KEY environment variable to enable AI features.

OPENAI_API_KEY="sk-..."

To route requests through a proxy or use an OpenAI-compatible provider (Azure OpenAI, Cloudflare AI Gateway), set OPENAI_BASE_URL:

OPENAI_BASE_URL="https://your-endpoint.com/v1"

Add both variables to your .env file and restart Quackback. AI features activate automatically once a valid key is detected.

Post summaries

Posts with comments display an AI-generated summary card on the detail page. Summaries help you understand what users are asking for without reading every comment.

Each summary includes:

  • Overview -- 1-3 sentences capturing the core user need
  • Key quotes -- Notable comments from users (not team replies)
  • Next steps -- Suggested actions like "Investigate...", "Reproduce...", or "Respond to..."

Staleness detection

Summaries detect when new comments arrive after the last generation. A Stale badge appears next to the timestamp, and the summary refreshes automatically in the background.

Summaries are generated and refreshed automatically. No configuration needed -- they appear on any post with comments.

Duplicate detection

When new feedback arrives, Quackback searches for existing posts that cover the same topic. The detection pipeline works in three stages:

  1. Vector similarity -- Compares embeddings of the new content against all existing post embeddings
  2. Full-text search -- Runs a keyword search to catch matches that embeddings might miss
  3. LLM verification -- Sends candidate pairs to the model to filter false positives and explain why posts are similar

High-confidence matches appear as merge suggestions in the triage view. You also see similar posts in the Merge section of any post detail page, with similarity scores and one-click merge.

Duplicate detection improves over time as your post library grows. The more posts with embeddings, the better the similarity search performs.

AI suggestions

Feedback from integration sources like Slack is analyzed and turned into actionable suggestions. Quackback runs each message through a quality gate, extracts themes, and matches them against existing posts.

You get three types of suggestions:

  • Merge posts -- Two existing posts that appear to be duplicates
  • Vote on post -- Incoming feedback that matches an existing post
  • Create post -- Feedback that doesn't match anything, with an AI-generated title

Review suggestions from Admin → Feedback → Incoming. Accept or dismiss each one with a single click.

See Triage AI suggestions for the full guide on reviewing, filtering, and managing suggestions.

Backfill embeddings

Existing posts created before AI was enabled don't have embeddings. Run the backfill command to generate them:

bun run ai:backfill

This processes all posts that lack embeddings and generates vector representations for each one. The command is safe to run multiple times -- it skips posts that already have embeddings.

Backfilling makes API calls for every post without an embedding. For large workspaces, this may take several minutes and consume API credits.

Privacy considerations

When AI features are enabled, Quackback sends the following data to your configured AI provider:

  • Post titles and descriptions -- used for embedding generation and duplicate detection
  • Comment text -- used for summary generation (only user comments, not team replies)
  • Candidate post pairs -- sent for LLM verification during duplicate detection

No user emails, names, or account metadata are included in AI requests. All requests go to the endpoint specified by OPENAI_BASE_URL (or api.openai.com by default).

If your organization has data residency requirements, use OPENAI_BASE_URL to route requests through a compliant proxy or provider. Review your AI provider's data retention policies before enabling AI features.

To disable AI features entirely, remove the OPENAI_API_KEY variable from your environment and restart Quackback.

Next steps