transcript.txt
user: How do I set up a retrieval pipeline with Anthropic Claude? assistant: You'll want to use the Messages API with a system prompt that instructs the model to extract structured knowledge from the conversation context. user: What about chunking strategies for long documents? assistant: For documents over ~50k tokens, use semantic chunking with overlapping windows. For shorter texts, a simple character-count threshold works. user: Can I visualize the extracted graph in the browser? assistant: Yes — use Cytoscape.js with the fcose layout extension for an organic force-directed visualization. Size nodes by degree. user: Claude, extract the graph from this transcript.
uses extracts enables

Chat transcripts are a mess.
Turn them into a graph you can actually explore.

Paste an AI conversation. Get a Cytoscape knowledge graph. No database. No backend. No account.

How it works

01

Paste your transcript

Drop one or more chat transcripts into the textarea. Raw text, any format.

02

Add your API key

Paste your Anthropic API key into a field. Stored in memory only — never saved to disk or localStorage.

03

Generate graph

Claude extracts entities, concepts, and relationships into a structured JSON graph. Cytoscape renders it with an organic force-directed layout.

04

Explore

Click any node to highlight it and its neighbors. Search to find concepts. Re-run the layout. Export the graph.

What you get

Force-directed layout

Organic, readable graph layout powered by Cytoscape.js with the fcose extension. No messy hairballs.

Color-coded by type

Concept, decision, tool, person, topic — each node type has a distinct color. A legend makes it clear.

Size by degree

Central concepts get bigger nodes. Peripheral topics stay small. The graph tells you what matters.

Click to explore

Click any node. It highlights along with its direct neighbors. A side panel shows the label and all connections.

Search nodes

Type in the search box to instantly highlight matching nodes. Find any concept without scrolling the graph.

Re-run layout

Stuck in a bad arrangement? Hit the layout button to re-run fcose and get a fresh graph.

Built for you, Olga

Stack

Plain React + Vite. No backend. No database. One static deploy.

Security

Your API key lives in React state only. Never written to disk, localStorage, or any log. You control the key completely.

Extraction

Claude Sonnet 4 via the Anthropic Messages API. max_tokens: 4000. Pronoun resolution, deduplication, short labels — all handled by the model.

Open flags

A — Chunking: Will long transcripts exceed 4000 tokens? Consider a chunk-and-merge step for transcripts over ~80k chars.
B — Schema mode: Optional fixed-schema extraction with predefined node types and relations for more deterministic output.

Every AI conversation you have contains a hidden structure — entities, decisions, tools, relationships. Graphify makes it visible.

Start with one transcript. See what you've been missing.

Try it now →