PaperTrail

Structure

Layers are inferred from where files sit in the tree, not from reading the code. A project that names its directories unconventionally will read oddly here — open the file browser to check anything the diagram implies.

Languages

Share of indexed source by file size. Binary and vendored files are excluded.

Dependencies

Declared in the repository’s manifests at the indexed commit. A declared package is not proof it is used, and runtime dependencies are listed first.

Feature verification

• Built with Letta, Jina, Supabase, Next.js/React/TypeScript, shadcn, Tailwind, Claude Sonnet 4.5Verified

  Built With: claude, css, jina, letta, next.js, postgresql, react, shadcn, sonnet4.5, supabase, tailwind, typescript

  Claimed on Devposthigh confidence

  • papertrail/package.json:19 — Dependencies include @letta-ai/letta-client, @supabase/ssr, @supabase/supabase-js, next, react, typescript, tailwindcss
  • papertrail/lib/agents/letta.ts:134 — Agents are created with model: 'anthropic/claude-sonnet-4-5-20250929', matching the Sonnet 4.5 / Claude claim
  • papertrail/components.json:2 — shadcn schema config file present, confirming shadcn/ui usage
• Learning from feedback (upvote/downvote) to refine future suggestionsVerified

  Learns from feedback to refine future suggestions and become more personal; each vote updates the user model

  Claimed on Devposthigh confidence

  • papertrail/app/api/paper-feedback/route.ts:76 — On upvote/downvote, the paper_likes or paper_dislikes memory block on the user's Letta agent is fetched, appended with the paper id, and updated
  • papertrail/db/migrations/20251028_paper_feedback.sql:1 — paper_feedback table persists per-user, per-paper vote type
• Multi-agent supervisor/worker orchestration for paper searchVerified

  Multi-agent orchestration (Letta): supervisor agent generates targeted queries; workers fetch papers via Jina AI and evaluate in parallel

  Claimed on Devposthigh confidence

  • papertrail/lib/agents/letta.ts:277 — runSupervisorWorkerFlow creates a supervisor agent, generates search queries, spawns 1-4 worker agents via createWorker, distributes candidate papers across them, and evaluates in parallel with Promise.all
  • papertrail/lib/agents/letta.ts:244 — createWorker instantiates a Letta agent per worker to evaluate paper relevance
• On-demand paper reading and real-time Q&AVerified

  Reads specific papers on demand and answers questions in real time

  Claimed on Devposthigh confidence

  • papertrail/lib/agents/letta.ts:92 — runReaderFlow detects a paper link/arXiv id in the message, fetches full content via Jina, and sends it to a persistent Letta agent for summarization; subsequent messages to the same agentId let the user ask follow-up questions
  • papertrail/lib/agents/letta.ts:170 — chatWithAgent routes messages containing arXiv ids/URLs or reader keywords (summarize/analyze/explain/read) into Reader mode
• Paper search via web search / JinaVerified

  Searches the papers using Jina AI to fetch content

  Claimed on Devposthigh confidence

  • papertrail/lib/agents/letta.ts:257 — fetchJinaSummary builds a request to https://r.jina.ai/ to fetch paper HTML/content for arXiv IDs or URLs
  • papertrail/lib/agents/letta.ts:193 — Letta agents are created with tools: ['web_search'] used by the supervisor to find candidate papers
• Persistent deep memory profile (interests, dislikes, goals, expertise, explanation style)Verified

  Deep memory architecture (Letta): profiles store interests, dislikes, goals, expertise, explanation style; conversation history; and feedback tables linking votes to preferences

  Claimed on Devposthigh confidence

  • papertrail/db/migrations/20251026_onboarding_profile.sql:4 — profiles table has expertise_level and explanation_style columns in addition to interests/dislikes/goals from the init migration
  • papertrail/lib/agents/letta.ts:220 — Supervisor agent is created with memory blocks (human, persona, preferences, paper_interests, paper_likes, paper_dislikes, skip_rules, reading_history, learned) mirroring the claimed deep memory architecture
  • papertrail/db/migrations/20251028_paper_feedback.sql:1 — paper_feedback table links user votes to paper ids, supporting the feedback-to-preference linkage claim
• Personalized paper recommendations returned to userVerified

  Returns relevant, personalized recommendations

  Claimed on Devposthigh confidence

  • papertrail/lib/agents/letta.ts:415 — Scores are sorted descending, top 6 selected, enriched with arXiv metadata, and returned as recommendations from runSupervisorWorkerFlow
  • papertrail/app/api/chat/route.ts:61 — Recommendations are persisted to conversation_messages and returned to the client
• Reading queue managementVerified

  Curate personalized academic paper recommendations and manage a reading queue (implied by dashboard/reading-queue UI)

  Claimed on Devposthigh confidence

  • papertrail/app/api/reading-queue/route.ts:32 — Full CRUD API for a per-user reading_queue table (add, list unread, mark read, delete)
  • papertrail/components/reading-queue.tsx:1 — Dedicated UI component for the reading queue
• Personalized relevance scoring against interests/dislikes/goalsCode-supported

  Scores paper relevance against your interests, goals, expertise, and feedback

  Claimed on Devpostmedium confidence

  • papertrail/lib/agents/letta.ts:365 — Worker evaluation prompt includes user interests and dislikes and asks for a 0-100 relevance score; localScore() provides a fallback heuristic using the same fields
  • papertrail/app/api/chat/route.ts:22 — expertise_level and explanation_style are fetched from the profiles table but the UserProfile type/object passed into chatWithAgent only carries interests/dislikes/goals, so expertise is not actually threaded into scoring, only partially matching the claim
• Sleeptime compute for generating insights across conversationsCode-supported

  Sleeptime compute generates insights over previous conversations, further personalizing the agent

  Claimed on Devpostmedium confidence

  • papertrail/lib/agents/letta.ts:136 — Agents are created with enableSleeptime: true, delegating to Letta's built-in sleeptime feature, and a 'learned' memory block is provisioned to hold generated insights
  • papertrail/app/api/agent/learned/route.ts:26 — Endpoint retrieves the 'learned' block value from the agent, presumably populated by Letta's sleeptime agent, but the actual insight-generation logic runs inside Letta's platform and is not implemented in this repo, so end-to-end behavior cannot be confirmed from code alone

An AI agent derived these features from the project’s Devpost page and readme, then searched the code for each one. Verified features are backed by cited code; claimed-only features had no supporting code, which is not by itself proof a feature is missing.