Turn scattered operational documents into grounded, conversational answers.
AI solution design · RAG implementation · conversational UX
Working prototype
RAG · Enterprise AI
Operational knowledge lives across PDFs, manuals, and engineering procedures. Finding the right passage requires searching multiple documents and judging which source is relevant.
A retrieval-augmented assistant brings document ingestion, semantic search, and cited responses into one conversational experience.
AI solution design · RAG implementation · conversational UX.
Indexing and answering are separate pipelines. Source metadata travels with retrieved passages into the response.
Source documents in S3 are read by a Node.js ingestion pipeline and split into approximately 500–700 token chunks. Amazon Titan Embeddings V2 creates vectors, which are stored alongside text and metadata in RDS PostgreSQL with pgvector.
At query time, the question is embedded with the same model. pgvector retrieves the five most relevant sections using an HNSW index. Source metadata stays attached to the retrieved context.
Claude Sonnet, accessed through AWS Bedrock, receives the retrieved passages and response instructions. The answer includes source references so users can inspect its evidence. Retrieval improves grounding; it does not guarantee correctness.
It keeps chunks, source metadata, and embeddings in one familiar database. A dedicated vector service may become useful at a different scale or with different operational requirements.
Approximate nearest-neighbor retrieval without an index training step. Memory use, recall, and index tuning remain tradeoffs.
The documented prototype uses this range to balance context with retrieval precision. It is a starting point to evaluate, not a universal optimum.
Working prototype documented in the original portfolio. The public demo uses fictional documents and preset responses. No production deployment or measured business impact is claimed.
Choose a sample question to explore a grounded response.
Preset responses for the sample topics shown above.
Run a sample to explore each step, the structured data, and the resulting action.
AWAITING SAMPLE INPUTInterested in enterprise AI, agentic systems,
or intelligent automation? Let’s connect.