DEV Community: Le Huy Hiep

We automated our Terraform. Locally it worked fine. In CI it hung for three hours.

Le Huy Hiep — Mon, 14 Sep 2026 14:36:09 +0000

It worked locally 😂 because Terraform asks for the variable it is missing, and something was there to answer. An agent, in our case, typing into the prompt.

In CI nothing types. Runner stdin never closes, so there is no timeout either. It just waits.

No error. Three hours of nothing.

Terraform wanted project_id. It was never set in CI.

Two fixes: pass it as TF_VAR_project_id, and add -input=false so it fails in two seconds instead of hanging.

The smaller cost was the three hours. The bigger one was deleting the plan gate on PRs. It hung at random and was not blocking anything, so it went.

userId! vs !userId

Le Huy Hiep — Mon, 07 Sep 2026 10:10:53 +0000

I never write the first one. But an agent could.

userId! tells TypeScript "trust me, this exists."
!userId checks whether it actually does.

My habit is the check: !userId, then act. The agent's output used
the assertion: userId!, assume it exists. On a route that allowed
guests, a user who was not logged in meant no user ID.

The code called the database with an empty ID. An empty ID means no
filter. The query returned every user's records.

No ID -> all the data.

My tests missed it because they only ran with logged-in users.
I caught it re-reading the code before a migration.

The fix: check for a missing ID and stop before the query runs.

When you review agent output, do not just check the logic. Check the
assumptions it typed as facts. The confident version is the dangerous
one.

Agent reading every files to answer one simple question?

Le Huy Hiep — Tue, 18 Aug 2026 05:30:57 +0000

My knowledge base grew across thousands of markdown files. Finding one fact meant reading whole files. Keyword search found words, not meaning. And reading whole files burns tokens: tens of thousands of tokens to surface one fact. The real metric is not storage. It is tokens per answer.

Why not the usual tools?

grep finds the file, but you still read it. Cost unchanged.
A cloud vector DB sends private notes out and adds infra for a 2 MB corpus.
Semantic-only search cannot do exact filters.

What I built: a local-first retrieval layer.

SQLite FTS5 index: instant keyword search, exact filters, real file:line references.
Local ONNX embeddings (384-dim): semantic ranking. Every hit points at a real file.
One-command rebuild. The index is derived, so it never goes stale.

Result: 2,000 chunks, one CLI, zero servers. An answer now costs a few hundred tokens, not a sweep through thousands of files.

Start deterministic. Add semantic where keywords fail. Rebuild often.

How to add UI for your RAG?!?

Le Huy Hiep — Mon, 03 Aug 2026 14:53:12 +0000

Just built some FastAPI SSE backend streaming LLM responses token-by-token,
after i got some free time.

Two things broke in production that worked fine locally 🙃:

nginx buffers SSE by default. The stream was delivering locally, dead
silent on Railway. proxy_buffering off, two hours later, fixed.
The LLM API returns an empty_retrieval event when no docs match. I wasn't
handling it, so the frontend sat on a loading spinner indefinitely. Added
the handler, resolved.

https://zooming-charm-production-547e.up.railway.app/