← Back to the vault

Lab Note · updated 2026-06-20

Why Fovea Lab Uses a Knowledge Graph

A recommendation is only as trustworthy as what it knows a tool can't do. Fovea represents tools, models, papers, and constraints as a graph with explicit negative edges — unsupported modalities, known failure modes, hard requirements — so the system can say "no" with a reason, not just rank positive matches.

knowledge graphcontrolled vocabularyretrieval
← Back to constellation

The thesis — A microscopy recommendation engine that only knows what tools can do will confidently recommend the wrong one. Ask it to segment an electron-microscopy image and, with only positive fit indexed, it will reach for Cellpose — because nothing in its knowledge says it shouldn't. The fix is structural, not a better prompt: represent every tool, model, paper, and workflow as a node in a graph whose edges include the negative constraints — unsupported_modalities, known_failure_modes, hard_requirements — so the system can decline with a reason. A flat list of capabilities can only ever say "yes." A graph with negative edges can say "no, because."

Why it matters — Three properties fall out of the graph that a vector blob of descriptions cannot give you. First, trustworthy exclusion: an entry that carries unsupported_modalities: [electron_microscopy] and hard_requirements: [GPU] lets the agent rule a tool out before recommending it, instead of discovering the mismatch in the lab. Second, controlled vocabulary: every tasks, modalities, and sample_types value resolves to a shared ontology, so "nuclei segmentation," "nucleus_segmentation," and "DAPI segmentation" are one node — the graph connects tools that a free-text search would scatter. Third, legible structure: the FoveaLab KB is itself a graph of ~1,100 nodes and ~5,800 edges clustered into 61 communities (Cell/Nucleus Segmentation, Deconvolution & PSF Analysis, Whole-Slide Imaging, …), so "what else is in this neighborhood, and what is known not to work here" is a traversal, not a guess. Each node also carries epistemic freshness (claims_valid_as_of, claim_expiry_risk) and a review tier, so unreviewed or stale claims never surface as confident advice.

How Fovea applies it — This is the model-isn't-the-system worldview made into data structure. The agent does not pattern-match a request to the most-cited tool; it traverses a graph where the negative space is first-class — what a tool refuses, what it requires, where it has been seen to fail. That is what lets a recommendation come with its own caveats attached, the same way a foundation model only means something wrapped in QC and correction, and the same reason validation insists on stating what was tested and on which population. The negative edge is the trust: a system that can articulate why it said no is one you can let near a real experiment. Otherwise you have rebuilt exactly the glue-and-configuration debt Sculley warned about — a powerful core surrounded by ungoverned assumptions.

A recommender that only indexes what tools *can* do will never tell you what *not* to use. Trust lives in the negative edges — unsupported modalities, failure modes, hard requirements — so model the "no," not just the "yes."

References

References

Local graph