Interesting project - it’s rare to see news-flow tracking done in real time at this scale.
One thing you may want to stress-test is how stable the clustering remains when stories evolve semantically over a few hours. Embeddings tend to drift as outlets rewrite or localize a piece, and HNSW can sometimes over-merge when the centroid shifts.
A trick that helped in a similar system I built was doing a second-pass “temporal coherence” check: if two articles are close in embedding space but far apart in publish time or share no common entities, keep them in adjacent clusters rather than forcing a merge. It reduced false positives significantly.
Also curious how you handle deduping syndicated content - AP/Reuters can dominate the embedding space unless you weight publisher identity or canonical URLs.
Overall, really nice work. The propagation timeline is especially useful.
A trick that helped in a similar system I built was doing a second-pass “temporal coherence” check: if two articles are close in embedding space but far apart in publish time or share no common entities, keep them in adjacent clusters rather than forcing a merge. It reduced false positives significantly.
Also curious how you handle deduping syndicated content - AP/Reuters can dominate the embedding space unless you weight publisher identity or canonical URLs.
Overall, really nice work. The propagation timeline is especially useful.