Insights are the structured knowledge Aira extracts from your sources. Each insight is a single, atomic piece of project intelligence backed by evidence from your data.
Atom kinds
Every insight has a kind that categorizes what it represents:
| Kind | What it means | Example |
|---|---|---|
| claim | A factual statement from the source material | "The API handles 10K requests/second in production" |
| decision | An architectural or product decision that was made | "We chose PostgreSQL over MongoDB for ACID compliance" |
| requirement | Something the project needs to satisfy | "Users must be able to export data as CSV" |
| risk | A potential problem or threat | "Single point of failure in the payment gateway" |
| unknown | Something the sources mention but don't resolve | "Authentication strategy not yet decided" |
| action_item | A concrete next step identified from the sources | "Need to schedule security audit before launch" |
| domain_signal | A market or domain observation relevant to the project | "Competitors are moving to server-side rendering" |
Searching and filtering
The insights page provides multiple ways to find what you need:
- Full-text search — Search across atom titles and bodies
- Kind filter — Show only requirements, or only risks, etc.
- Domain filter — Filter by the domain area (frontend, backend, infrastructure, etc.)
- Tag filter — Filter by tags extracted during analysis
- Entity filter — Filter by named entities (people, services, technologies)
- Include dismissed — Toggle to include previously dismissed insights in results
Search is powered by PostgreSQL full-text search indexes over both atoms and source chunks.
Evidence anchors and provenance
Every insight links back to its source evidence. Click an insight card to see:
- Evidence snippets — The exact text from the source that supports this insight
- Source reference — Which source document the evidence came from
- Chunk and line range — The specific location within the source
- Evidence hash — A cryptographic verification that the evidence hasn't been modified since extraction
This provenance chain means you can verify any insight against its original source. Nothing is fabricated — every atom traces back to evidence.
Dismissing vs deleting
- Dismiss — Hides the insight from default views and excludes it from synthesis (PRD generation, feature generation). The atom stays in the ledger. You can bring it back later with the resurrect action.
- Delete — Not supported for ledger atoms. The append-only design ensures auditability. Dismiss is the correct workflow.
Contradictions
When Aira detects two atoms that conflict with each other, it creates a contradiction record. For example:
Atom A: "The system uses JWT for authentication" Atom B: "Session-based auth is the current approach"
Contradictions appear with a severity level and require human resolution. You can:
- Resolve — Choose which atom is correct and provide a rationale
- Dismiss — Mark the contradiction as not meaningful
- Investigate — Leave it open for further analysis
The contradiction workflow is a core part of the human-in-the-loop (HITL) design. Aira surfaces conflicts; humans decide.
Confidence scores
Each insight carries a confidence score (0 to 1) based on:
- Number of evidence rows supporting it
- Directness of evidence (direct quotes score higher than inferred references)
- Cross-source reinforcement (same claim from multiple sources increases confidence)
- Kind-specific evidence floors (decisions and requirements need at least 2 evidence rows unless marked draft)
You can use confidence as a filter to focus on high-confidence insights or review low-confidence ones that may need additional sources.