Source top-up: highlight text → search primary sources — honest, optional, and without recalculating the analysis

The problem — and why we don't promise "everything verified"

AI analyses and chat answers can contain plausible-sounding but incorrect provisions, rulings or case numbers. That is not a SmartLegalPro bug — it is a limit of every language model.

What we do not want to do: slap "unverified" on everything across the board, or pretend we have checked every statement against the internet. That would be dishonest.

What we do instead: distinguish three clear states:

1. From the document — a quote or fact that appears in the original (analysis with source_quote).
2. AI assessment — model knowledge without a separate web search (standard chat and standard analysis).
3. Source search performed — you actively started "Add sources"; the result is a saved research log.

How the source top-up works

Where: the Vault (document analysis) and the document chat.

Flow:

1. Highlight at least 30 characters (a sentence, paragraph or bullet).
2. Toolbar "Add sources" — the price appears before you start (CostHint, action key source_verify_expand).
3. Confirmation dialog with a compliance note: research log, no guarantee.
4. Perplexity sonar-pro searches for primary sources (statutes, rulings, authorities) — only for the highlighted text, without recalculating the analysis (expand, not recalculate).
5. Every URL is fetched; quotes are checked against the target page:

- Badge "Quote found on page" — a snippet match on the live page.

- Badge "Quote not confirmed" — source found, wording not verifiable on the page (not automatically "wrong").

1. Claims without a primary source are listed separately.
2. The result is saved in source_expansions — viewable per document in the Vault overview.

Cache: the same text + country within 24 hours → result from cache, 0 Credits.

Monetisation — transparent and admin-configurable

• Action key: source_verify_expand.
• Default: 1 Credit.
• Admin range: 0–10 Credits (platform admin → Action Costs).
• Billing: only after confirmation; refund on a technical error.
• Standard chat: no web search included — only the chat_message price.

What really works — and what doesn't

Works well

• Targeted search for a single highlighted passage with context from the analysis/document.
• Structured sources with type, court/authority and case number only when it is stated on the source page.
• Snippet check against the fetched HTML page (no blind trust in the AI answer).
• Persistence: your personal research archive per document — the foundation for later Vault insights / wiki.
• PII masking before the Perplexity call (as in other flows).

Limits — please plan realistically

• No full verification: no one can guarantee "every ruling worldwide". The search is best-effort.
• The snippet check is heuristic: dynamic pages, paywalls or PDF-only rulings can miss matches → "not confirmed" ≠ "wrong".
• Perplexity errors: if the API fails, we refund your Credits — but you get no result.
• Jurisdiction: primarily the document's country; cross-border remains complex.
• No substitute for a lawyer: even with sources it stays information work, not advice.

Compliance & prompts

• Chat system prompt (chatAntiHallucinationBlock): no invented rulings, no false "source search performed" claim.
• Perplexity prompt (perplexityVerifyExpand.ts): verify-only, JSON schema, claims_without_source, hints on national primary sources.
• UI disclaimer on every result panel; FAQ platformFaq.webSearchExtra.
• Credit history: readable labels, including refunds.

Vault wiki — what's possible next

Every source top-up is already a structured row in the database (excerpt, summary, sources JSON, status, timestamp). That is the basis for:

• pinning / bookmarking in the Vault (planned),
• cross-references to the case extract and analysis,
• a later searchable document wiki built from your own research.

Today: history below the analysis + an inline panel after each search.

Where to try it

1. Vault → open a document with an analysis.
2. Highlight a passage → "Add sources".
3. Below the analysis: expand Source top-ups (n).

The same flow in the document chat for a linked document.

Feedback right on the result (👍/👎) helps us improve prompts and domain filters.