Telecom Standards Assistant: 3GPP RAG

During 5G deployments at Rogers, I watched engineers spend hours digging through dense 3GPP specifications to answer basic architecture questions. A single query could span hundreds of pages across multiple documents (TS 38.300 plus 38.401 plus 23.501), all written for compliance rather than for reading. Keyword search in PDFs has low recall on conceptual questions, and standards work needs citations an engineer can trust inside a time-sensitive change window.

• 5G RAN engineer: ask in plain English, get a cited answer in seconds, and resolve implementation questions in minutes instead of hours.
• Standards researcher: query across releases at once to track how a feature evolved across spec versions.
• New engineer: ask the "dumb questions" about 5G architecture without blocking a senior, in a multi-turn chat that remembers context.
• Grad student: use it for research with no API subscription, because the whole thing runs free and local.

• Semantic search with a source citation and similarity score on every answer.
• Domain and generation filters (5G/LTE, RAN/Core) to scope a query to the right spec subset.
• Token-by-token streaming and multi-turn conversation memory.
• 37 specs, 43,121 chunks, fully local on ChromaDB + Ollama at $0 per query.
• A public Streamlit demo so anyone can try it in the browser.

The public demo reached 25+ users in its first three months. On a held-out evaluation set it scores 100% context precision, 80% context recall, and ~50ms p50 retrieve latency, backed by 152 fully-mocked unit tests. Getting it public also meant clearing two security-hardening rounds (42 findings: rate-limiting gaps, prompt injection, SSRF, missing headers), which I wrote up on Substack.