The problem is larger than generating an answer
A useful news reader needs more than a prompt and a chat model. It must acquire material from changing sources, preserve links and dates, make content available in several languages, and retrieve relevant items later. Each stage can fail independently, and an answer is only as useful as the records behind it.
Wassla therefore separates preparation from reading. It processes an article when it is ingested; a later reader request can use the stored title, summary, translation, and vector instead of repeating every model call. This reduces query-time work, but it means the collection can be stale between ingest runs and every stored transformation still needs evaluation.
Follow the data, not the library names
The database has two roles: it stores sources and articles, and it starts the periodic ingest request. The pipeline service fetches RSS, runs the article workflow, writes results, and exposes read and ingest endpoints. The Nuxt web service asks for read data and renders pages. Phoenix receives traces so a developer can inspect what happened inside the pipeline.
The communication paths are deliberately different. Postgres calls the pipeline over the private Compose network for scheduled ingestion; the pipeline calls feeds and model APIs; the web server calls the pipeline through a read-only proxy; and traces flow from the pipeline to Phoenix. The browser does not receive the ingest token or call the ingest endpoint.
RSS feeds -> FastAPI and LangGraph -> PostgreSQL and pgvector
| |
+-> model APIs +-> scheduled ingest request
+-> Phoenix traces
Browser <- server-rendered Nuxt <- GET/HEAD proxy <- pipeline APIWhy use separate services?
One alternative is to fetch and translate an item only when a reader opens it. That keeps storage smaller, but makes page latency depend on several external calls and repeats work for every reader. Another is a single process that serves the interface and runs ingestion. That is simpler at first, but mixes public read traffic with a privileged write operation and makes it harder to test each boundary.
Wassla chooses four local Compose services: Postgres, the pipeline, the web app, and Phoenix. This makes the service contracts visible and gives a new developer one command to start the system. It is a reproducible development topology, not evidence of high availability: one Compose host, one database volume, and external model providers remain important dependencies.
Reproduce the baseline
Clone the repository, copy .env.example to .env, set a private INGEST_TOKEN, and choose either configured model credentials or the deterministic stub providers. Then run make setup and make up. Stub mode lets the graph and interface run without paid model calls; its translated text and hash-based vectors are test aids, not meaningful language or search quality.
The next note starts at the first external boundary: how the scheduled request fetches RSS, rejects unsuitable entries, and chooses which new items enter the article graph.
cp .env.example .env
# Set INGEST_TOKEN; for a smoke run, set both providers to "stub".
make setup
make up