Skip to content

“Search got worse as the documents piled up.”

Document search that gets worse as documents pile up

More documents mean more near-misses, slower queries and an index someone has to look after. DocSlurp runs the whole search stack, so volume is our problem, not your cluster’s.

Why it happens

  • Vector search alone returns plausible neighbors, not the right passage, once a corpus holds many similar documents.
  • One shared index makes every collection compete with every other for the top results.
  • When ingestion runs on the machines that answer queries, a large upload slows search for everyone.

What changes with DocSlurp

  • Hybrid search, reranked

    Keyword and vector results are combined, then reranked. Every hit carries the passage, its file and page, and a relevance from 0 to 1.

    Details
  • An index per project

    Each project and embedding model gets its own vector database, so collections never share an index.

  • Ingestion off your servers

    Parsing, OCR and embedding run on DocSlurp’s workers. Keyword search covers a document as soon as it is chunked, and each response says which parts are ready.

  • Tune one request, not the system

    Raise the candidate depth to 200 or turn reranking off for one query; project defaults stay as they are.

    Details

search.ts

for await (const hit of client.search({ q: 'termination notice period' })) {
  console.log(hit.relevance, hit.citation.filename, hit.citation.pageNumbers);
}

Sound like your setup?

Tell us what you run today and what it costs. We’ll say plainly whether DocSlurp fits.

Jump to