Context & Retrieval: Handling Information
Activity Overview
Modern models have massive context windows (up to 2 million tokens). This activity investigates how models handle large amounts of data and external information, and where the limits of "retrieval" lie.
Long-context Q&A
The Needle in the Haystack
Upload a long text (PDF) or provide a 1-hour YouTube video link to a model that supports it (like Gemini 1.5 Pro).
Ask detailed questions about specific timestamps or page numbers to test retrieval accuracy.
"What did the speaker say at minute 42?"
"Summarize the argument made on page 15."
Retrieval Toggle Comparison
To Search or Not To Search
Ask a question about a recent event (e.g., "What was the closing stock price of NVDA yesterday?").
- Case A: Search/Grounding ON.
- Case B: Search OFF (Explicitly prompt: "Do not search the web").
This tests the model's knowledge cutoff vs. its ability to use tools.
Context Window Explorer
Lost in the Middle?
Use the Model Context Explorer tool (or manually inject a "needle" fact in the middle of a long document).
Ask the model to retrieve that fact. See if placement (beginning, middle, end) changes the reliability of the answer.