Browse categories
Explore
Fiverr Pro
English
$
USD
I will tune pinecone qdrant or opensearch for rag
ali shah
About this gig
Bad answers in a RAG app are usually a retrieval problem, not a model problem. I configure, tune and debug your vector database Pinecone, Qdrant or OpenSearch so the right chunks reach the prompt.
WHAT I DO
- Schema & ingestion: embedding dimension, similarity metric, HNSW parameters, clean upsert/delete pipelines in Python or TypeScript
- SDK wiring: LangChain, LlamaIndex or raw client, with retries and idempotent indexing
- Retrieval quality: metadata filtering, multi-tenant namespace isolation, top-k and score thresholds, hybrid BM25 + vector, reranking
- Pinecone: serverless vs pod, namespaces per tenant, index cost tuning
- Qdrant: cloud or self-hosted Docker, payload schema, quantization, snapshot strategy
- OpenSearch: k-NN on your existing cluster without losing keyword search
- Debug & migrate: empty or irrelevant results, latency, runaway spend, ChromaPinecone/Qdrant migrations verified with query suites
If chunking is the real cause, I show you the evidence instead of invoicing you for index tuning. No chat UI, no prompt engineering, no fine-tuning.
Before ordering, send: DB, embedding model, vector count and one failing query.
Get to know ali shah
ali shah
AI LLM Engineer RAG, Agents, LLM Ops Shopify Hydrogen
- FromPakistan
- Member sinceAug 2026
Languages
Urdu, English
5 years shipping production software (Systems Limited, then Contour
Software) as a backend + AI engineer. I build the unglamorous parts that
make AI products work: RAG retrieval that actually returns the right
chunk, evals that catch hallucinations before customers do, LiteLLM
gateways and Langfuse traces so you know your per-user cost, and vLLM on
your own GPU. Also custom Shopify Hydrogen storefronts.
.
My Portfolio
Other AI Development Services I Offer
FAQ
Pinecone, Qdrant or OpenSearch — which should I pick?
Pinecone if you want managed and zero ops and accept per-unit cost. Qdrant for control, self-hosting and strong filtering. OpenSearch if you already run it and want hybrid keyword + vector in one place. I'll argue for whichever fits, including "stay put".
My RAG returns irrelevant chunks. Is that your gig or the fix gig?
If it's index schema, filtering, top-k or embeddings, it's this gig. If it's prompt logic, agent routing or model behaviour, that's my debugging gig. Send one failing query and I'll route you correctly, even when that costs me the sale.
Can you migrate Chroma or Weaviate to Pinecone?
: Yes — rebuild schema and ingest pipeline, then dual-write, backfill, and verify with a query suite before you cut over. You keep the old index until parity is proven, so a migration never becomes an outage.
Can you fix my vector DB bill?
Often. Cost usually comes from oversized dimensions, per-tenant indexes instead of namespaces, unbounded deletes leaving orphan vectors, or over-fetching top-k. I audit read/write patterns and give you the config change plus expected savings.
Do you do hybrid search and reranking?
Yes — BM25 + vector fusion where supported, and cross-encoder reranking when your corpus warrants the latency. I benchmark before and after so you can see the accuracy delta rather than take my word for it.
