Research material has a life after the report. An interview can contain observations that become relevant to a different question. First, however, someone needs to find that material and understand the context in which it was produced.
I built an interview archive with metadata, vector search and MCP access. It provides ways to reach accumulated research material through search and an agent interface. The public demonstration uses synthetic interviews; real material requires authentication.
Material and context
A transcript lets a researcher return to the conversation. Metadata helps them orient themselves within the archive. These layers do different jobs: the text contains the material, while its description helps establish what the material is.
Vector search adds a route through semantic similarity. That can be a useful starting point when a new question uses different words from the conversation it needs. Similarity alone, however, does not establish whether a retrieved passage supports a new research conclusion.
For me, retrieval and interpretation belong together. Finding a passage becomes useful when there is also an opportunity to examine where it came from and what it means in context. A search result is the beginning of further work with evidence.
Access through an agent
MCP connects the archive to an agent interface. A question can be asked in the environment where work is already happening, with the retrieved material available for the next research step. My contribution was building the archive, its metadata and vector search, and this way of accessing it.
There is an important distinction between obtaining relevant material and answering the research question. An agent can help reach the archive; the argument still depends on examining what has been found.
What is public
The Atlas demo presents the public demonstration using synthetic data. The demo agent connection page provides a way to explore agent access within the same demonstration boundary.
These examples introduce the interface and the service’s underlying idea. They do not expose real interviews. Separating the two makes it possible to show the tool while keeping public examples distinct from the team’s research material.