Study Compares 500- and 1500-Token Chunks for Marketplace Handbook Retrieval

·4 views

A DEV Community article examines how chunk size affects retrieval quality when answering questions from a marketplace product handbook. It argues that 500-token chunks can isolate a relevant clause but drop its heading or exceptions, while 1500-token chunks keep more context but may bury the matching rule among unrelated text. The author recommends splitting along heading-bounded sections first and comparing fixed sizes against that baseline using the same questions and citation requirements. The piece also stresses tracking document revision and chunk identity so citations resolve to the revision used, and proposes a 30-question evaluation fixture covering direct lookups, cross-section questions, and unanswerable cases. The article treats the problem as one of grounding answers in verifiable evidence rather than simply tuning chunk size.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

FastAPI CSV Upload Service Had Four Bugs Its Tests Missed

The author found that a CSV upload endpoint returned a plain 500 error instead of the intended 413 response when a file exceeded the size limit. The root cause was a nonexistent status constant name in the Starlette version used, which raised an AttributeError that the service's error handling did not catch. The service also lacked an encoding policy, so Latin-1 files and binary files renamed as CSVs were handled without validation. A size check that ran only after the entire file was read, along with a more permissive-than-expected CORS configuration, were also identified as problems. The author fixed each issue and added targeted tests, including one that checks rejection of oversized files by temporarily lowering the limit.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Daily Payment Reconciliation Scheduling Relies on Idempotent Run Keys and Webhook Retries

A DEV Community write-up argues that scheduler choice should be based on the delivery guarantees it can actually prove, not on the convenience of its interface. For a healthtech payment reconciliation job, the author recommends at-least-once execution paired with an idempotent report key. The webhook is treated only as a trigger, while an immutable run record serves as the authoritative source of truth. The article proposes scheduling a run at 02:00 UTC, with separate boundaries tracked for scheduler delivery, run claiming, report publishing, and consumer acknowledgement. It also explains that an idempotency key such as the business date, rather than an "exactly once" label, lets repeated webhook deliveries map to a single logical run.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Recurring HTTP Triggers Can Drive Bounded Data Cleanup in B2B Systems

The article proposes using an external recurring scheduler to send a small, authenticated HTTP request to a cleanup endpoint instead of running an internal task executor. The endpoint claims a bounded time window and delegates the actual processing to a separate worker pool, keeping the trigger stateless. Shorter intervals reduce cleanup latency but raise request pressure, so the endpoint should cap concurrency, honor Retry-After headers, and make retries safe. The author suggests a starting policy of a five-minute tick with a one-minute window and a two-minute safety delay for clock skew and late writes. Responses such as 202 Accepted or 429 Too Many Requests are recommended to signal accurate status to callers.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Debugging DNS Resolution Failures When the Server Itself Is Reachable

A server can answer pings and still be unreachable by name because DNS resolution is failing, which is often the true cause of the problem. The author outlines a practical troubleshooting sequence, beginning with dig and nslookup to determine whether the resolver or the remote host is at fault. Common culprits include an overwritten /etc/resolv.conf pointing to a dead nameserver, an unreliable internal resolver, systemd-resolved stub confusion, stale or poisoned caches, and firewall rules that block UDP port 53. To prevent recurrence, the author recommends pinning a known-good resolver configuration in systemd-resolved and monitoring DNS query latency alongside ping checks. The full postmortem with packet captures is published on the author's ops notes site.

0 comments Read more at DEV Community