RAG system

IC at Genpact

OpsDesigner

Structured as a LangChain-based retrieval-first system where source quality, document shaping, and answer transparency matter as much as generation quality.

Built as an individual contributor at Genpact using LangChain, focused on retrieval quality, grounding, and maintainable tuning workflows.

LangChainRetrieval pipelinesSemantic searchChunk strategyPrompt engineeringFeedback loops

System flow

LangChain RAG retrieval flow

Shape

A retrieval-first workflow where query shaping, source selection, context assembly, and answer generation stay inspectable and tunable.

Tap a handoff to inspect that part of the flow.

Current handoff

Shape

User question to Query shaping

The system reframes the user question into a retrieval-ready intent before searching.

Overview

What this system does

OpsDesigner is a LangChain-based RAG assistant built around grounded operational knowledge. The goal is to help users ask practical questions and receive answers that are tied back to retrievable context instead of generic model output.

The system treats retrieval quality as a product problem, not just a vector search setting. Query shaping, chunk strategy, context assembly, answer framing, and feedback loops all affect whether the final answer feels trustworthy.

Contribution

My role in the build

The parts of the system where my ownership and design work mattered most.

01

Built as an individual contributor, focused on retrieval flow, LangChain composition, answer grounding, and maintainable tuning paths.

02

Worked on shaping the RAG workflow so retrieval, context assembly, prompt behavior, and response framing could be improved independently.

03

Focused on making the system easier to tune after launch by keeping retrieval behavior and answer quality visible.

Architecture decisions

Design choices that made the system easier to reason about

Decision 01

Treat retrieval as a workflow

The system separates question shaping, retrieval, context building, generation, and review so relevance issues can be isolated.

Decision 02

Keep grounding visible

Answers are designed around source-aware framing so users can understand where the response is coming from.

Decision 03

Make chunking tunable

Document structure and chunk strategy are treated as design decisions because they directly affect retrieval precision.

Decision 04

Route feedback back into retrieval

User and quality feedback should improve retrieval settings, prompt layers, and knowledge operations rather than disappear after the response.

Reliability

How trust was designed into the workflow

Source-aware answer framing to reduce unsupported responses.

Feedback loops for identifying weak retrieval results or poor answer framing.

Evaluation hooks for tracking retrieval relevance and response quality.

Separation between retrieval configuration, context assembly, and final generation so tuning stays manageable.

Workflow behavior

How the system moves from input to output

01

A user asks an operational question.

02

The system shapes the question into a retrieval-ready intent.

03

LangChain retrieval components search the knowledge layer.

04

Relevant context is assembled and passed into the response layer.

05

The answer is generated with grounding in mind, then feedback routes into future retrieval tuning.

Tech stack

Technology grouped by responsibility

Retrieval

LangChainSemantic searchChunk strategy

Grounding

Context assemblySource-aware promptsAnswer framing

Quality loop

Feedback captureEvaluation hooksRelevance tuning

Next improvements

What I would strengthen next

Add a more formal retrieval evaluation set with representative operational questions.

Introduce reranking and query expansion experiments for harder questions.

Track answer confidence, retrieval misses, and source coverage as observable quality signals.