Core Lens

Overview

Core Lens is an agentic AI research platform built to close the gap between “an assistant that sounds right” and “an assistant you can check.” It plans multi-step research across the open web and a user’s own library of uploaded papers, then grounds every answer in sources you can click through to — paper title, section, and page number included.

Key Features

• Agentic research tools: A toolset the model can call on its own — web search, academic search, library search over uploaded papers, deep multi-step “extreme search”, PDF search scoped to an attached document, live URL retrieval, and translation.
• Personal research library: Upload and organize academic PDFs into tagged collections, run full-text search across everything, and pull specific papers into a conversation with @mentions to cross-reference and synthesize.
• Traceable citations: Answers carry inline citations with source tracking, so a claim can always be traced back to the passage that produced it.
• 90+ models, one interface: Model routing across xAI, OpenAI, Anthropic, Google, DeepSeek, Qwen, Mistral, Cohere, GLM, MiniMax and more — many with both standard and reasoning modes — behind a single, consistent chat surface.

Technologies Used

• Next.js + Vercel AI SDK: Streaming responses, tool calling, and the agent loop.
• Vercel AI Gateway: Multi-provider model routing without provider-specific plumbing in the app.
• Exa AI: Web and academic search, plus content retrieval for live crawling.
• PostgreSQL + Drizzle ORM: Conversations, library metadata, and citation records.
• Redis / Upstash: Caching and rate limiting.
• Better Auth: GitHub, Google, and X OAuth.
• Tailwind CSS + shadcn/ui: Interface layer.
• Docker: Multi-stage Node 22 Alpine build with Compose for local and production parity.

Challenges and Learnings

The hard part was never generating text — it was keeping the chain from question to source intact. Retrieval had to preserve enough structure (section, page, offset) to render a citation a researcher would actually trust, which meant treating chunking and metadata as first-class rather than an implementation detail of the vector store. Supporting 90+ models across a dozen providers also forced a strict separation between the agent’s tool contract and each provider’s quirks, so adding a model became configuration rather than code.

Outcome

Core Lens turns a scattered research workflow — tabs, PDFs, half-remembered papers — into a single searchable surface where the AI’s reasoning is auditable. It’s the clearest expression of the thing I keep coming back to in my work: AI output is only as useful as your ability to verify it.