Ingest accepts unbounded documents and inserts one chunk per round trip

[royalpinto007]

Problem

POST /documents accepts a document of unbounded size and indexes it one round trip at a time.

DocIn in app/main.py puts no limit on text:

class DocIn(BaseModel):
    id: str
    title: str
    source: str
    text: str

Compare AskRequest.question, which is carefully bounded: Field(min_length=1, max_length=2000). The bound was clearly considered on one side of the API and not the other.

Then ingest in app/ingest.py does:

vectors = embedder.embed([c.text for c in chunks])   # every chunk in one call
...
for c, vec in zip(chunks, vectors):
    await conn.execute("INSERT INTO chunks ...", (...))   # one round trip per chunk

A 20 MB document becomes roughly 17,000 chunks. Those are embedded in a single SentenceTransformer.encode call, so the full result set is materialized in memory at once, and then inserted with 17,000 sequential round trips inside a single transaction that holds locks on the document's rows the whole time. doc_acl inserts have the same one-at-a-time pattern in a loop.

Since the endpoint currently has no authentication (#5), this is reachable by anyone who can reach the service.

Suggested approach

1. Bound the input. Add max_length to DocIn.text and title, and a maximum chunk count in ingest, returning a 413 with a clear message rather than trying and dying. Pick numbers deliberately and put them in app/config.py next to the other settings.
2. Batch the embedding. Feed embedder.embed in fixed-size batches instead of one call per document, so memory is bounded by batch size rather than by document size.
3. Batch the inserts. psycopg 3 has cursor.executemany with pipeline support, and COPY for the larger case. Either turns thousands of round trips into a handful.
4. Same treatment for the doc_acl insert loop.
5. Add a test that ingesting a document near the limit stays within a reasonable number of database round trips. Counting round trips is a fair proxy and is stable in CI, unlike timing.

Done when

• Oversized documents are rejected with a clear error rather than accepted.
• Chunk and ACL inserts do not scale as one round trip per row.
• Embedding memory is bounded by a batch size constant, not by document size.
• The limits are configurable and documented in .env.example and the README.

---

If you want to take this on, comment on the issue to claim it and it will be assigned. Please keep to a maximum of 2 open claims per person at a time so other contributors get a chance.