Mission archive chatbot / Interactive demo · in progress

Apollo 11 Mission Archive

Explore the landing site in 3D and ask questions about Apollo 11, with answers linked to mission documents.

My contribution: I built the document retrieval pipeline, chat interface and connection between scene selections and mission questions.
Tools & methods: Three.js · TypeScript · PDF retrieval · BM25 · Vector search

Interactive demo · Explore the scene and ask the mission archive.

Screenshot of the running Apollo 11 chatbot and 3D scene. Open the interactive demo to orbit the scene or ask a question. Scene assets retain their original credits in the demo.

The problem

What this project investigates

Mission details are spread across long reports and transcripts. Finding a relevant passage is only part of the problem: an answer needs to preserve its context and let the reader inspect the source.

The build

What I built

A browser-based lunar scene connects to a mission-document chatbot. Selecting an astronaut, the lunar module or an experiment changes the question context. The backend retrieves passages from indexed PDFs and streams an answer with links to source pages. The scene uses credited third-party models; it is a stylised illustration, not a surveyed reconstruction of the landing site.

Technical choices

Decisions behind the build

Combine keyword and semantic retrieval

BM25 keyword search and cosine similarity retrieve complementary candidates. Reciprocal rank fusion combines their rankings before passages are passed to the answer model.

Keep source pages within reach

PDF page citations let a reader check the original material rather than relying on the generated answer alone.

Use scene selections as context

Selectable objects connect the visual scene to relevant questions. Each submitted question is answered independently; the visible transcript is not conversational memory for the model.

Current status

What the evidence shows

The implementation includes the scene, retrieval service and streaming chat. Evaluation notes record failures involving numeric tables, event timing and unsupported premises. A targeted table guard addresses some numeric questions, but does not establish general answer reliability. This is an experiment in progress, not a verified mission reference service.

Further work & evidence gaps

Further evaluation is needed for ambiguous event times, flattened PDF tables and questions with incorrect assumptions. The public demo uses a hosted retrieval service and model API. Answers still need to be checked against the cited documents.