An LLM assistant reachable over plain SMS, built for areas of the Philippines with slow, intermittent, or no internet access. Text a question, get an answer back as a text — no app, no data connection required.
A text arrives through Semaphore (a Philippine SMS
gateway) and hits a webhook Lambda on AWS. The Lambda embeds the question,
scores it against a DynamoDB table of embedded content (weather, PAGASA
cyclone advisories, news, and a curated document corpus), and passes the
best-matching chunks to Amazon Nova Lite via Bedrock's converse API. For
general questions outside that corpus, the model answers from its own
knowledge; for time-sensitive topics (news, prices, weather, schedules,
current officials, sports), it's restricted to what retrieval actually
found, so a stale answer never gets presented as current. The reply is
trimmed to fit a single SMS segment and sent back through Semaphore.
A second, scheduled Lambda keeps the time-sensitive content fresh by re-fetching weather, PAGASA advisories, and news feeds on an hourly cadence and re-embedding them into the same table.
Full architecture, request flow, and design rationale: docs/system-design.md. Product goals, constraints, and cost tradeoffs: docs/context.md.
- Python, deployed as two AWS Lambda functions with zero third-party dependencies (everything AWS goes through boto3)
- Amazon Bedrock: Nova Lite for generation, Cohere Embed Multilingual for embeddings
- DynamoDB for the vector store (brute-force cosine similarity, no ANN index — the corpus is a few hundred chunks) and for per-number rate limiting
- Semaphore for inbound/outbound SMS
Every component is pay-per-request; there's no always-on infrastructure. See Cost posture in docs/context.md for why that mattered enough to shape the architecture.
src/handler.py— the webhook Lambda: parses the inbound SMS, rate-limits the sender, retrieves and generates an answer, sends the reply.src/ingest.py— the scheduled Lambda that keeps weather, PAGASA, and news content current in the vector store.scripts/ingest_corpus.py— local-only tool for embedding a curated document corpus; not deployed.template.yaml— AWS SAM template describing the architecture. Reference only; the actual deployment is done by hand in the AWS Console (seedocs/deployment.md).tests/— pytest suite; Bedrock and DynamoDB are faked, no AWS access needed to run it.docs/— product context, system design, and the deployment runbook.
Requires a Python virtual environment at .venv.
.venv\Scripts\python -m pip install -r requirements-dev.txt
.venv\Scripts\python -m pytest tests
Deployment is manual, console-only — no CLI or CloudFormation apply. Steps are in docs/deployment.md.