Now with agentic web fallback

Your documents.Finally intelligent.

Cortex turns your PDFs, documents, and notes into a smart, conversational knowledge base — hybrid search, AI re-ranking, and source citations on every answer.

Start for Free

See How It Works

app.cortex.ai/dashboard

Workspace

Annual_Report_2024.pdf

Q3_Results.pdf

Strategy_2025.pdf

HR_Policy_v3.pdf

Upload document

Ask your documents

What was our Q3 revenue and how does it compare to last year?

Based on Q3_Results.pdf, Q3 revenue reached $4.2M — up 23% YoY from $3.4M in Q3 2023, driven by enterprise contract growth in APAC.

Q3_Results.pdf · p.12

Annual_Report_2024.pdf · p.4

Ask a question about your documents…

Architecture

From question to cited answer — six real steps.

This is the exact retrieval pipeline running behind every chat message — not a diagram made for marketing.

STEP 01

Ingest & Chunk

Uploaded PDFs and documents are parsed and split into overlapping chunks, ready for embedding.

document_chunks

STEP 02

Matryoshka Embeddings

Each chunk is embedded with gemini-embedding-001 and truncated to 768 dimensions, then stored in Supabase pgvector.

768-dim

STEP 03

Hybrid Retrieval

pgvector cosine similarity and Postgres full-text search run in parallel, fused with Reciprocal Rank Fusion.

RRF · k=60

STEP 04

AI Re-ranking

The top 10 candidate chunks are scored for relevance and cut down to the 3 that actually answer the question.

gemini-3.1-flash-lite

STEP 05

Agentic Decision

The agent checks whether the retrieved context is sufficient, or reaches out to live web search when it isn't.

tool-calling · Tavily

STEP 06

Streamed Response

The final answer streams back token-by-token over Server-Sent Events, cited to the exact source chunk.

SSE · gemini-2.5-flash

Under the hood

Built for correctness, not vibes

Every answer is retrieved, re-ranked, and streamed by real infrastructure — here's what's actually running.

Retrieval

Finds what you mean, not just what you type

Every query runs two searches in parallel — dense vector similarity and Postgres full-text search — then fuses the results with Reciprocal Rank Fusion.

vector cosine+BM25 full-textRRF · k=60

Storage

768-dim, workspace-isolated

Matryoshka-truncated embeddings stored in Supabase pgvector, scoped per workspace so documents never cross tenants.

Re-ranking

10 → 3 candidates

The top 10 retrieved chunks are scored for relevance and cut down to the 3 that actually answer the question.

Streaming

Answers arrive as they're generated

Q3revenuereached$4.2M—up23%YoYfromenterprisecontractgrowthinAPAC.

Agent

Knows when to look further

When your documents don't have the answer, the agent calls out to live web search instead of guessing.

Cited, always

Every claim traces back to the exact document and chunk it came from — click a citation to see the source.

Q3_Results.pdf · p.12Annual_Report_2024.pdf · p.4

Your knowledge base is one upload away.

Drop in your PDFs and start asking questions in seconds. No setup, no config — just answers.

Get Started — It's Free

No setup · No config · Just upload and ask

Cortex

Get Started