Build your own Perplexity.

# Build an original app inspired by Perplexity

You are an experienced product designer and full-stack engineer. Build a working, polished MVP from the brief below. Implement the core journeys end to end; a static mockup is not sufficient. Inspect the current environment, make sensible implementation decisions, and get started. Ask a question only if a missing answer genuinely blocks implementation.

Starting platform: Web. Explicit customized Product & audience and Tech stack requirements take precedence over this starting target.

## Product & audience
Build an original AI answer-and-research app called Sourcepath, inspired by Perplexity. A user asks a question, the server retrieves results from a configured search service, and a configured model produces an answer tied to those sources. Follow-up questions retain a bounded research thread. This wraps real search/model providers and does not recreate their indexes or models.

## Design
Use an original search-first interface with a focused query box, visible searching/generating states, a readable answer, and expandable source cards. Clicking a citation should reveal the exact supporting source. Distinguish verified source references from model commentary and partial/failed answers.

## Features
1. Real retrieval: query one configured search API with limits and preserve result URLs, titles, snippets, and retrieval timestamps. If full-page retrieval is supported, fetch only public HTTP(S) resources through a bounded safe fetcher.

2. Grounded answers: pass a bounded source set to a real model, require source IDs for factual claims, and render citations only when they resolve to retrieved records. Reject invented citation IDs and show insufficient-evidence states.

3. Follow-up threads: save private question/answer history and referenced source snapshots. Resolve follow-up context explicitly without mixing another user’s research or allowing unbounded context growth.

4. Streaming and controls: stream partial answers, allow cancellation, and retain completed/failed/interrupted states. Display provider setup errors, quotas, and source retrieval failures rather than fabricated search results.

5. Export and privacy: copy or export an answer with source URLs and timestamps. Rate-limit queries, cap spending/concurrency, redact credentials in logs, and let users delete their research history.

## Tech stack
Use a TypeScript frontend, Python or TypeScript API, PostgreSQL, and current documented search/model provider APIs with server-only keys. Treat retrieved pages as untrusted evidence, not instructions. If fetching URLs, deny private/reserved networks at DNS resolution and every redirect, cap sizes/time, and avoid authenticated sites.

## Data & behavior
ResearchThread {ownerId}; Turn {query, state, answer, providerUsage}; SearchRun; Source {url, title, snippet, retrievedAt, contentHash}; Citation {turnId, sourceId, marker}; RetrievalFailure; UsageWindow. Store enough source metadata to explain an answer without pretending search snippets are full-page verification.

Acceptance checks and scope:
Test real retrieval and generation, a claim with no source, fabricated citation IDs, contradictory sources, follow-up isolation, cancellation, search failure, and attempted internal-network retrieval. Exclude model training, unrestricted crawling, paywall bypass, guaranteed factual correctness, and autonomous actions based on retrieved text.

## Originality and assets
The named product is a reference for general workflows only. Use an original product name, visual identity, wording, layout, and assets. Do not copy proprietary code, branded logos, distinctive screens, private APIs, paid templates, or protected media. Use original or appropriately licensed sample data and assets. Build from this brief rather than scraping the reference service. Do not imply affiliation with the referenced company.

## Implementation quality
- Make every visible control work, with clear loading, success, empty, validation, and failure states. Remove controls for unimplemented features.
- Support the target platform and responsive or resizable layouts where appropriate. Use platform-appropriate accessibility semantics, labels, keyboard operation, visible focus, sufficient contrast, and reduced-motion preferences. Provide alternatives for drag-only interactions.
- Validate inputs at the relevant boundary. Treat imported or user-generated content as untrusted; render it safely. Confirm destructive actions or offer undo.
- Keep state changes reliable across navigation and reloads. Never silently lose edits. Handle persistence and network errors with an actionable message.
- Keep the scope realistic. Prioritize the complete primary journey. If a custom Features section differs from the baseline, update the data model, design, and acceptance checks to match it; explicit custom requirements take precedence over baseline assumptions.
- Provide realistic original sample content and a useful empty state. Keep sample data clearly separate from user-created data.

## Delivery and verification
1. Inspect the repository and outline the implementation in a few sentences.
2. Implement the app and connect its actual data flow. Continue until the primary journey works.
3. Run appropriate checks and test the acceptance cases above on the target platform. Verify responsive layouts and keyboard navigation when browser tools are available. For native projects, build with the available SDK and test an emulator or device when possible; report any missing native verification honestly. For games, verify the rules and gameplay loop, not only the menus.
4. Fix observed failures, then provide the runnable result, setup/run instructions, a brief explanation of what is implemented, and any specific configuration or verification limits.

This is a focused MVP. Do not add billing, AI services, or unrelated features unless explicitly requested in the brief. Start building now.