Ada

Background

After Sponge (document pipeline) I wanted to query knowledge contextually, not just index it. Enterprise programme leadership generates much distributed context — Ada is the attempt to make it locally usable.

Problem

Programme and transformation work generates scattered notes, architecture documents and conference material — querying that context quickly matters, especially when external AI services are not an option.

Goals

• Local RAG pipeline without cloud dependency
• Ingestion for Markdown and PDF
• Source citations in every answer
• Evaluate chunking strategies for heterogeneous documents

Architecture

Technology stack

TypeScript · Electron · Local LLM · Vector DB · RAG

Engineering challenges

• Chunking strategies for heterogeneous documents
• Latency of local models on consumer hardware
• Making source citations in answers traceable
• Treating setup complexity as a product decision

Trade-offs

• Local models: privacy vs. latency on consumer hardware
• Electron: fast UI vs. resource consumption
• RAG without MCP tooling — manual integration instead of standardised tool contracts

Lessons learned

• Local AI is feasible, but setup and model choice are product decisions
• RAG quality depends more on ingestion than on the model
• Personal use is the best test for operations problems

Current status

Live (Private) — executive intelligence for opportunities, developments, and long-term knowledge.

Future improvements

• Refine chunking strategies (PDF vs. Markdown)
• Evaluate MCP tool layer — explicit tool contracts for local actions
• Observability for RAG pipeline: which chunks, which model, which latency

Estimated milestones

• Q2 2026 — Chunking evaluation with representative documents
• Q3 2026 — MCP prototype for local tool calls
• Q4 2026 — RAG observability dashboard