Timeline
14 weeks
Team
5 engineers + 1 PM
Client Stage
Enterprise Pharma Consulting
The Problem
Pharma consultants drowned in scattered data — publications, internal reports, databases — with no time to find credible, contextual answers. Existing AI and BI tools returned shallow results without citations, forcing users to re-verify everything manually.
Our Approach
Data Audit & Source Prioritization (Week 1–3)
Catalogued internal reports, licensed databases, and public publications. Defined citation standards and the evaluation rubric for answer quality (accuracy + citation traceability).
Ingestion & Vector Pipeline (Week 4–6)
Built ingestion workflows that respected licensing terms, chunking strategies tuned for biomedical content, and Milvus vector stores for semantic retrieval with metadata filters.
Agentic RAG with LangGraph (Week 7–11)
Implemented multi-step agent flows: plan → retrieve → synthesize → cite. Agents can drill into sub-questions autonomously while surfacing their reasoning trace for the consultant to verify.
Citation UX & Rollout (Week 12–14)
Designed the citation UI so every claim in every answer links back to the source paragraph. Rolled out to the full consulting team with onboarding workshops.
The Solution
We built a conversational research assistant powered by agentic RAG. Users ask questions in natural language and receive cited, contextual answers drawn from curated pharma datasets and documents. The system performs deep dives into publications, links new findings with prior studies, and recommends related resources — every response backed by transparent source references.
Why This Tech Stack
LangGraph was chosen over vanilla LangChain for the explicit graph-based agent control we needed for multi-step research plans. Milvus (over Pinecone) because the client had strict data residency requirements and wanted self-hosted vector infrastructure.
The Outcome
Transformed how the consulting team conducts research. Queries that previously took hours of database searching now return cited answers in seconds, and the transparency of sources gave the team confidence to rely on AI-generated insights in client deliverables.
Key Metrics
2 hrs → 8 sec
Research query time
> 97%
Citation accuracy
+2.1× per week
Consultant deliverable throughput
> 90% in 60 days
Adoption across consulting team