BFSI workloads · Case Study 02

Enterprise Agentic RAG

I kept metadata and vectors together in PostgreSQL/pgvector. LangGraph handles planning, retrieval, evaluation, and answer synthesis so the flow stays auditable.

PRODUCTIONBack to case studies

status

PRODUCTION

environment

GCP Kubernetes (GKE)

ingress

Istio Ingress Gateway

runtime graph

6 nodes / 6 edges

View system map

System map

Enterprise Agentic RAG

Env: GCP Kubernetes (GKE)Ingress: Istio Ingress Gateway

Problem: I designed an agentic RAG pattern for long financial filings. The system needed layout-aware parsing, hybrid retrieval, clear state transitions, and a grounding check before any answer could ship.

My engineering note

I kept metadata and vectors together in PostgreSQL/pgvector. LangGraph handles planning, retrieval, evaluation, and answer synthesis so the flow stays auditable.

Live path

request: Istio gateway -> FastAPI parser

Path running

Failure modes

Mode

Readout:LangGraph is routing the query. pgvector is serving grounded candidates.

Signals

Retrieval

pgvector

Hybrid metadata and vector lookup in the relational boundary.

Orchestration

LangGraph

Stateful retrieve, evaluate, and synthesize nodes.

Runtime

GKE/Istio

Containerized service boundary with cluster ingress.

Reliability

Eval gate

Grounding and safety checks before response finalization.

Run logs

> GKE: RAG service received governed request through Istio ingress.

> PARSER: Filing sections normalized with source coordinates and metadata filters.

> PLANNER: LangGraph state advanced from classify_query to retrieve_context.

> PGVECTOR: HNSW candidate set merged with relational metadata constraints.

> EVAL: Grounding check compares answer claims with retrieved evidence.

> TRACE: Run ID, retrieved chunk IDs, and evaluation decision persisted.

Architecture Decision

Why I chose this design.

Short decision notes tied to the code or config that mattered.

Decision

rag_graph.py

I used LangGraph with PostgreSQL/pgvector because the RAG flow needed auditable states and retrieval metadata near the vector index.

Code noteSee engineering notes

rag_graph.pypython

python

from langgraph.graph import END, START, StateGraphfrom langgraph.checkpoint.postgres import PostgresSaver
class RAGState(TypedDict):    query: str    filters: dict[str, str]    retrieved_docs: list[Document]    grounding_score: float
def compile_rag_graph(pool):    graph = StateGraph(RAGState)    graph.add_node("plan", plan_query)    graph.add_node("retrieve", retrieve_from_pgvector)    graph.add_node("evaluate", evaluate_grounding)    graph.add_edge(START, "plan")    graph.add_edge("plan", "retrieve")    graph.add_edge("retrieve", "evaluate")    graph.add_edge("evaluate", END)    return graph.compile(checkpointer=PostgresSaver(pool))

from langgraph.graph import END, START, StateGraphfrom langgraph.checkpoint.postgres import PostgresSaver
class RAGState(TypedDict):    query: str    filters: dict[str, str]    retrieved_docs: list[Document]    grounding_score: float
def compile_rag_graph(pool):    graph = StateGraph(RAGState)    graph.add_node("plan", plan_query)    graph.add_node("retrieve", retrieve_from_pgvector)    graph.add_node("evaluate", evaluate_grounding)    graph.add_edge(START, "plan")    graph.add_edge("plan", "retrieve")    graph.add_edge("retrieve", "evaluate")    graph.add_edge("evaluate", END)    return graph.compile(checkpointer=PostgresSaver(pool))

Next case study

GPU Platform Modernization

Inference scaling - I worked on inference platform patterns where static GPU allocation slowed teams down. Production serving needed quota, ...

Need this level of architecture review?

Bring the hard system constraint: retrieval quality, agent failure modes, latency, evaluation, deployment topology, or technical market education.

Start Advisory Intake View GitHub