[[Graphify]] is a derived layer behind whatever the organisation already uses to author knowledge — the question is whether the extracted graph carries enough of [[Collective knowledge management|collective knowledge management]]'s requirements to justify the migration effort and the ongoing [[LLM]] cost. ![[Is Graphify a good fit for personal knowledge management?-2026-05-27T154452.png|400]] ## [[Graphify]] ![[Graphify#^definition]] Scoped to the documents side - prose, PDFs, transcripts, images - the code pass is set aside. ## [[Collective knowledge management]] fit Each requirement below gets a verdict against [[Graphify]] on the documents side. Each sub-section transcludes the underlying pain question and follows with Graphify's actual fit - what holds, what is qualified, what is weak, what is not addressed at all. ### [[Discoverability]] ![[Knowledge problems#^discoverability]] Mixed fit, degrading at scale. For modest corpora the god-nodes report and graph traversal surface concepts that a flat folder leaves invisible. But the god-nodes report is bounded by what a human can scan: as corpora grow, the most-connected topics produce nodes whose neighbourhoods are themselves too large to read. Compositional topics make this worse - a god node for "Knowledge" encompassing thousands of related concepts collapses into noise. Findability holds for small clusters and mid-fan-out nodes; it degrades exactly where it matters most, on the largest and most central topics. ^discoverability ### [[Connectivity]] ![[Knowledge problems#^connectivity]] Mixed fit. Graphify does land edges across document formats - PDFs, transcripts, prose, images - in one graph without prior schema work. Two issues qualify the fit: First, every edge carries a confidence tier (`EXTRACTED`, `INFERRED`, `AMBIGUOUS`). The edges are *claims*, not facts. An edge at confidence 0.6 is a different object than a hand-authored "A depends on B", and downstream consumers must either threshold or weight the edges - they cannot treat the graph as a set of asserted relationships. Second, the human's authoring role over edges is unclear: #foresters/research Can a human add an edge, annotate an existing one, or override a confidence tier in Graphify's output? If edges are read-only output of the extraction pipeline (regenerated each run), the "graph of meaning" is one-way: machine-authored, with the human as observer. Connectivity without authoring authority is connectivity in shape only. ^connectivity ### [[Agent-readiness]] ![[Collective knowledge problems#^agent-readiness]] Fit depends on the agent's task. Structurally the output is strong: JSON in NetworkX node-link format, queryable via CLI and an MCP server, no reformatting layer between the graph and the agent. Semantically it is weaker: the agent cannot treat edges as facts, because each carries a confidence tier and may be wrong. An agent doing *research* can [[Reasoning|reason]] about confidence, selectively traverse uncertain edges, and surface candidates for human review - the probabilistic graph is a fit. An agent doing *fact retrieval* for an authoritative answer cannot trust an edge without further validation - the same graph is a no-go. The split holds at scale: the larger the corpus, the more confidence tiers there are to reason over. ^agent-readiness ### [[Maintainability]] ![[Knowledge problems#^maintainability]] Weak. Incremental updates only link a new document to nodes it explicitly names or to other files in the same extraction batch, so cross-corpus semantic similarity drifts as new documents arrive. Restoring it requires a full rebuild that pays the [[LLM]] cost on the whole corpus, and that cost grows linearly with corpus size rather than with what changed. ^maintainability ### [[Survivability]] ![[Knowledge problems#^survivability]] Weak in isolation. The source documents survive in the editing tool, but the graph itself is derived - and because the schema is emergent, a rebuild on the same corpus may produce a different relation vocabulary. The graph artifact is reproducible in shape but not in detail; long-term continuity of the *knowledge structure* depends on storing the artifact, not just the inputs. ^survivability ### [[Ownership clarity]] ![[Collective knowledge problems#^ownership-clarity]] Weak. Every edge on the docs side is LLM-produced and tagged `EXTRACTED`, `INFERRED`, or `AMBIGUOUS`, with no human author, reviewer, or retirer. The emergent schema is itself unowned - it changes between runs without anyone deciding it should. Organisations needing audit trails or attributable edits will find the output difficult to govern as a system of record. ^ownership-clarity ### [[Securability]] ![[Collective knowledge problems#^securability]] Not addressed. Graphify inherits whatever access boundaries the source documents already have; the graph artifact is a single file that either is shared or is not. Granular per-node or per-relation permissions require an external consumption layer. ^securability ## Verdict - [/] [[#Discoverability]] - [/] [[#Connectivity]] - [/] [[#Agent-readiness]] - [-] [[#Maintainability]] - [-] [[#Survivability]] - [-] [[#Ownership clarity]] - [-] [[#Securability]] ^verdict Adopt where research-style discovery is the dominant need - mixed formats, confidence-aware agents, modest corpora. Avoid where [[Ownership clarity|attribution]], governance, or [[Securability|access control]] matter: the probabilistic edges and emergent schema that make discovery cheap also make these unenforceable. ^answer The conclusion lives in none of the requirements on its own. Every `[/]` partial in the verdict is an axis that looked strong before a scale or confidence caveat re-priced it; the answer emerges only when those caveats are weighted against the organisation's actual scale and use case.