clusterComments is the whole pipeline: embedMany, candidatePairs, evaluatePairs, clusterLinks, then names. The same clusterLinks call runs over raw cosine similarity to produce the embeddings-only baseline. Embedding, Jev model, and naming are injectable so tests never touch the network.

Pure average-linkage clustering over compared pairs only. An uncompared pair is unknown, not zero. A merge needs two compared pairs unless only one pair exists, and validateCohesion ejects members under the floor. Do not replace this with connected components: one loose chain would merge unrelated themes.

candidatePairs keeps a pair when either comment ranks the other in its six nearest neighbors. This bounds what Jev is asked so pair count grows with the collection, not its square.

evaluatePairs asks one Jev boolean per candidate pair in batches of 40 with bounded concurrency and zeroDataRetention. A pair with no usable probability stays uncompared and is never treated as a match.

buildPairQuestions and THEME_FLOOR (0.50). The same-theme criteria define the granularity of the themes. Change the criteria, not the clustering code, to make themes broader or narrower.

Theme names are enrichment from generateText with reasoning off. cleanLabel trims model text to one short plain line and a failed name falls back to Theme N. Names never change membership.

clusterComments is the whole pipeline: embedMany, candidatePairs, evaluatePairs, clusterLinks, then names. The same clusterLinks call runs over raw cosine similarity to produce the embeddings-only baseline. Embedding, Jev model, and naming are injectable so tests never touch the network.

Pure average-linkage clustering over compared pairs only. An uncompared pair is unknown, not zero. A merge needs two compared pairs unless only one pair exists, and validateCohesion ejects members under the floor. Do not replace this with connected components: one loose chain would merge unrelated themes.

candidatePairs keeps a pair when either comment ranks the other in its six nearest neighbors. This bounds what Jev is asked so pair count grows with the collection, not its square.

evaluatePairs asks one Jev boolean per candidate pair in batches of 40 with bounded concurrency and zeroDataRetention. A pair with no usable probability stays uncompared and is never treated as a match.

buildPairQuestions and THEME_FLOOR (0.50). The same-theme criteria define the granularity of the themes. Change the criteria, not the clustering code, to make themes broader or narrower.

Theme names are enrichment from generateText with reasoning off. cleanLabel trims model text to one short plain line and a failed name falls back to Theme N. Names never change membership.