Browse categories
Explore
Fiverr Pro
English
$
USD
I will build a production rag pipeline over your business knowledge
Ehsan
About this gig
RAG demos are easy. RAG that survives real questions is not.
I build retrieval pipelines that hold up when a real user asks something awkward: hybrid retrieval (vector plus keyword), a reranking stage, chunking designed around your document shape rather than a fixed token count, and an evaluation set so you can prove accuracy instead of hoping for it.
What you get:
- Ingestion for your real sources: PDFs, docs, sheets, websites, database tables, support tickets
- Cleaning and one canonical document representation, the step most pipelines skip
- Hybrid retrieval plus cross-encoder reranking, tuned against your own questions
- A golden question set with scored results, so a regression is visible before your users find it
- Grounded generation with citations back to the source chunk
- Deployed as an API you own, or straight into your existing app
I run this on a production multi-tenant system, including the unglamorous parts: re-indexing when content changes, cost per query, latency budgets and failover lanes.
Tell me what your documents are and what people need to ask them.
Get to know Ehsan
Ehsan
AI Agent and RAG Engineer
- FromIndia
- Member sinceJun 2024
- Avg. response time1 hour
Languages
English
I build AI agents that answer from your business, not from guesswork.
My focus is retrieval - RAG: getting your catalogue, docs and policies into a pipeline the agent can be held to, so it cites real sources and escalates what it cannot ground.
I run a multi-tenant agent platform in production, so I work on the unglamorous parts daily: re-indexing, evals, cost per query, human handoff.
Typical work: WhatsApp and website support agents, internal knowledge assistants, custom tool-using agents, and audits of RAG systems that fail in production.
I say so when an agent is the wrong tool.
FAQ
Which vector database do you use?
Postgres with pgvector by default: cheap, and one less system to run. Pinecone, Qdrant or Weaviate if you already use them or prefer them.
Which LLM will it run on?
Yours or mine. OpenAI, Anthropic, Gemini, or open-weight models you host. Self-hosting genuinely gets cheaper past a certain volume and I will tell you honestly where that line falls for your usage.
Can it run entirely on our own infrastructure?
Yes, including self-hosted models, with no document content leaving your environment. Tell me your data-residency constraints up front and I will scope to them.
How do I know the retrieval is actually accurate?
You get a golden question set with scored retrieval results on your own data. Accuracy is reported with numbers, not claimed. On Standard and Premium you also get before and after scores.
Can you work with scanned documents or images?
Yes, with OCR. Flag it when you message me so I scope it properly, because scanned sources change the ingestion work significantly.
