Everyone can spin up a vector search in a notebook. Making it work as an actual product — across many sources, ranked usefully, fast enough to feel instant — is a different sport. That gap is where I spent my time building the GovernAI Research Atlas.
From "embeddings demo" to "discovery product"
The toy version of semantic search is three lines: embed some text, store the vectors, query for the nearest. The product version has to answer harder questions. Where do the vectors live so search stays fast at scale? How do you keep the index fresh as sources update? How do you unify wildly different content types into one coherent search? None of that shows up in the demo.
The Atlas unifies papers, repositories, and governance resources into a single semantic search experience — which means all of those questions had to be answered for real.
The stack that makes it work
- ChromaDB for vector search. A purpose-built vector store handles the one job that gets brutal at scale: finding the nearest vectors among many, fast. Doing that with a brute-force scan works in a notebook and dies in production; a real vector database is what keeps retrieval interactive.
- Sentence-Transformer embeddings. The quality of your search is capped by the quality of your embeddings — they're what turns text into the meaning-space where "closest" corresponds to "most related." Choosing and applying them well is upstream of everything else.
- Multiple sources, one space. Content pulled from sources like OpenAlex and GitHub is embedded into a shared space so a single query can cross formats — papers, code, and policy — instead of forcing three separate searches.
- A custom relevance score on top. Nearest-neighbor is the retrieval; a scoring layer decides what actually surfaces first, because the closest vector and the most useful result aren't always the same thing.
The takeaway
Vector search is easy to prototype and genuinely hard to productionize — the difficulty is all in scale, freshness, source diversity, and turning "similar" into "relevant." The tooling (a real vector DB, good embeddings) matters, but so does the engineering discipline around it.
Building the Atlas taught me where the notebook ends and the product begins. The full retrieval architecture is on the project page.
👉 Explore it: www.divyakush.com/projects/governai-research-atlas
Divyakush Punjabi — Full-Stack & AI Systems Engineer
🌐 https://www.divyakush.com · 💼 LinkedIn · 💻 GitHub
Top comments (0)