Zekher

Mission

Zekher (זכר, "remembrance" in Hebrew) is a project that integrates high-quality sources about the Holocaust into the generative artificial intelligence answer pipeline. The goal is to preserve Holocaust memory with AI—both by mitigating Holocaust denial in large language models' answers and by connecting users with verified information and survivor testimonies.

Problem

Even before users encountered Holocaust denial content from xAI's Grok, a UNESCO report warned of risks to Holocaust memory both embedded in the technology and possible with bad actors. LLMs have invented survivor quotes and “hallucinated” about less-documented history. But generative AI is so ubiquitous that not confronting these challenges risks mass-misinformation.

Resources

Zekher processes, stores, and utilizes documents from Yad Vashem's Holocaust Lexicon and Dr. David P. Boder's survivor interviews. There is no chunking or AI summary as part of the ingestion process, just PDF to text conversion and cleanup. Post-ingestion, Zekher’s technology does not strain museum or university servers.

That information powers a familiar Chat experience. The system uses Lexicon content to answer user’s queries before offering real interview audio snippets as a supplement. The system is agentic, meaning that the LLM has access to tools that query the data sources and enable citations. Zekher’s system prompt tells the model to acknowledge if information is missing in the Lexicon and prohibits summarizing survivor testimony. Users can read and listen to the sources once the model sees them.

Zekher provides a Model Context Protocol (MCP) server that can be used to add Lexicon information into Anthropic's Claude, OpenAI's ChatGPT, etc. But it could perhaps be more useful for developers of other agentic chatbots who would like to prevent Holocaust denial in their applications.

All sources are displayed in a unified library that is the landing page for citations from both Chat and MCP.

Acknowledgements

Thank you to the many people who have advised and supported this project in 2023. Every conversation has led to a feature in V3. Most importantly, Zekher now answers questions with updated information rather then data supplied by survivor testimony. This both avoids outdated information and misrepresenting survivor testimony.

Zekher was designed to be an open-source project. The first version is still hosted on GitHub, and this version will be added once any vulnerabilities are patched. This project would not be possible without other open source projects. You can find them here.