A working semantic index for personal storage — the foundation for a future conversational librarian.
I designed Harbor's indexing layer to make a large personal file collection understandable by meaning, while keeping the original folder structure intact.
Watch it unfold
harbor-preindex building the semantic map used by Harbor.The problem
Large personal storage trees stop being searchable once you no longer remember exact filenames or folder paths. A useful system must search by meaning, reduce the decision space, and stay transparent enough for a human to trust it.
Design & implementation
The running component, harbor-preindex, crawls the storage tree, turns each folder into a text profile, embeds that profile locally, and stores the result in Qdrant.
Working today
- Local indexing of the existing folder tree
- Folder descriptions and embeddings stored in Qdrant
Planned next
- Natural-language retrieval, such as "Where is my plumber invoice from last year?"
- Suggested destinations for incoming files, with human review before filing
The planned incoming area keeps files awaiting review separate from the durable collection. Conversational retrieval and assisted filing are not yet implemented.
Why I built it
I want finding a file to start with what someone remembers about it, not its exact path. The index is the foundation for that interaction.
