Local retrieval for difficult recall
Real-time recall under pressure is a hard problem; a searchable, source-linked record is an easy one. This tool converts the first problem into the second — locally, with nothing leaving the machine.
§1 The situation
Interview preparation runs on recall, and recall under pressure is exactly when unstructured memory is least dependable. The raw material existed — curated evidence, worked answers, skills, role responsibilities, spread across Markdown, JSON, and documents — but career material is private, which rules out shipping it to an external service to make it searchable.
§2 What I did
Implemented a local Python builder and query CLI: it creates a SQLite database, indexes the curated sources with FTS5, normalizes queries, ranks matches, falls back to broader search when a query comes up empty, and prints source-linked results. Broad Markdown ingestion deliberately excludes private and generated directories, and the database is disposable by design — rebuildable from sources at any time. The work is recorded as AI-assisted.
§3 Result
A rebuildable, network-independent way to retrieve interview evidence, curated answers, skills, and role responsibilities — each result linked back to the file it came from. This is the rare plate marked verified: the implementation was inspected and behaves as described.
§4 Difficulty & judgment
The design tensions were small but real: retrieval recall against understandable output, broad ingestion against private directories that must never be indexed, a persistent index against the freedom to throw the database away. Each one resolved the same direction — toward the version that is easier to trust.
§5 Not claimed
Deliberately left unclaimed
- The truth of every indexed claim — "verified" covers the inspected implementation behavior, not the content the tool indexes.
- The indexed material itself — the private career content stays private; only the tool is described here.
- Sole authorship — the implementation is AI-assisted, and labeled that way.
§6 The same work, other lenses
- Accessibility
- Reduced dependence on real-time unstructured recall by providing searchable, source-linked preparation material.
- Engineering
- SQLite schema creation, FTS5 indexing, ingestion, query normalization, ranked search, and fallback behavior, implemented in Python.
- Business systems
- Scattered interview evidence turned into a rebuildable local retrieval workflow with source paths and predictable commands.