Private · local-first knowledge systems · Graph tooling
Investigation Workbench
A local-first workspace for turning fragmented source material into reviewed, queryable knowledge without letting automated ingestion silently become authoritative truth.
- Niche
- Evidence-heavy knowledge review and graph exploration
- Core stack
- Memgraph · Python · Obsidian · Docker · MCP
- Model
- Raw source → candidate → reviewed canonical knowledge → compiled graph
- Focus
- Provenance, reviewability, temporal structure, reproducible projection
The niche
Graph and AI tooling becomes dangerous when extraction, interpretation and truth are treated as the same operation.
This workbench is aimed at material where chronology, relationships and source provenance matter enough that an automated system should propose structure without being allowed to silently finalize it.
Raw documents enter a controlled ingestion layer. Candidate knowledge is reviewed before promotion. Only reviewed canonical notes are compiled into graph, timeline and viewer artifacts.
What is different
The product treats review state as architecture, not workflow etiquette.
Ingestion can propose entities, events and relationships, but downstream graph truth is derived only from reviewed canonical material.
The core workspace runs locally with explicit storage, container and resource boundaries rather than assuming a cloud service should own sensitive source material.
Graph nodes, edges, timelines, manifests and viewer artifacts are compiled from the reviewed layer so outputs can be rebuilt and checked.
Obsidian, Memgraph Lab, generated timeline/graph viewers and MCP-connected agents consume projections of the same reviewed knowledge rather than inventing parallel truths.
Product & design thinking
The central product problem is epistemic: make it obvious what is source evidence, what is machine suggestion, what has been reviewed, and what is derived.
That distinction shapes navigation, file structure, validation and graph semantics. The system is designed to make uncertainty and provenance visible instead of producing an authoritative-looking graph from opaque extraction.
The same reasoning guides AI integration. Agents and MCP tools are useful as query and transformation interfaces, but their outputs remain bounded by explicit schemas, review gates and deterministic compilation steps.
Development approach
The repository is organized as an inspectable pipeline rather than a monolithic AI application.
Docker isolates Memgraph and its tooling; Python dependencies are managed reproducibly; the knowledge vault carries schemas and controlled vocabularies; validation covers links, temporal semantics and orphaned material; fixture cases exercise chronology and concurrency before real data is introduced.
The system also exposes MCP entry points so different AI clients can query the same local graph without each becoming a separate data authority.
Private repository · public case study intentionally omits case-specific source material