Expertise · RAG and knowledge systems

Reliable answers from your own knowledge

Retrieval-augmented generation lets a language model answer from your documents, with the source attached. The principle is simple. The quality is in the details, and that's where the developers in our network excel.

Where the quality of a RAG system comes from

  • Document processing. Tables, scans and long contracts each need their own approach. Badly chunked documents produce bad answers.
  • Search, not just generation. Hybrid search (keywords plus meaning) and reranking decide whether the model gets the right passages.
  • Permissions and freshness. Users only get answers from what they are allowed to see, and new documents are searchable within minutes.
  • Measured on real questions. An evaluation set of real questions and good answers runs with every change.
What that looks like in code
rag/search.pyPython 3.12
from pydantic import BaseModel

class Hit(BaseModel):
    doc_id: str
    text: str
    score: float

async def search(query: str, user: User, k: int = 8) -> list[Hit]:
    # only documents this user is allowed to see
    allowed = await acl.documents_for(user)
    # keywords and meaning combined, then reranked
    hits = await index.hybrid(query, doc_ids=allowed, k=k * 4)
    return reranker.top(query, hits, k=k)

Request profiles

Profiles of suitable RAG engineers within 48 hours

  • No obligation: you only pay for hours worked
  • Dutch-speaking, from our own network and assessed by us
  • On-site, hybrid or remote

Indicative € 106–132 per hour, all-in, excl. VAT

Rather call? +31 35 7111 373

What we select for

  1. 01

    Knows where errors come from

    A bad answer usually comes from bad retrieval. The developers in our network start improving there.

  2. 02

    Works with evaluation sets

    Improvements are measured, not felt.

  3. 03

    Privacy by design

    Knows which data may and may not go to an external model, and how to enforce that technically.

Technology we often see

Tools change fast. We select on principles, but this is what the developers in our network use in practice.

PythonpgvectorElasticsearchQdrantLlamaIndexUnstructuredCohere RerankRagas

Profiles we are looking for

Need a Python developer for this?

Send us your vacancy or request. Within 48 hours you hear which developers in our network fit and what it costs.