UI: send-on-Enter + live reasoning trace; docs: answer-cache decision (#106)

cleanup: remove the translation module — the guard is English-only now (#105)

app: live bare spinner under the answer during the freshness check (#103)

freshness: heal drifted chunks completely — re-embed in place, not deferred (#102)

freshness: stale-while-revalidate — self-heal cited pages, correct the live answer on drift (#101)

retrieval: remove the cross-encoder reranker (ablation showed no earned keep) (#100)

app: animate the thinking indicator so the answer wait never looks frozen (#99)

ci: Backfill content workflow (skip crawl) + empty-snapshot purge guard (#98)

perf: store chunk content in Neon so hydration needs no live fetch (Stage A) (#97)

perf: hydrate retrieved sections concurrently — cut the dominant answer latency (#96)

Follow client-side redirects + read_page URL guard: real API content (#92)

Make the deploy usable: strip/redact API keys + gate non-English + faster QuOTE (#91)

Concise answers + immediate thinking indicator (Ultralytics-style UX) (#90)

Route questions: grounded first, loop for multi-source shapes (#89)

QuOTE questions in the enrichment pipeline + one-click eval default (#88)

Reranker as the precision stage + fold glossing into Build Index (#85)

Answer-quality eval (LLM-judge) + hy3-led OpenRouter default (#81)

Fix synopsis extraction: read the Sphinx definition line, not the signature head

Extractive synopsis: lead every api chunk with the page's own summary sentence

Widen the HNSW candidate scan: kind-filtered dense queries were dropping true neighbors

Harden the OpenRouter path: no-choices 200s retry, dead deepseek slug removed

Contextual Retrieval: LLM glosses for api reference pages

Fix per-pool relevance gap: a close tutorial no longer empties the api pool

Add the agentic benchmark: coverage via citations, agentic vs single-shot

Add the v1 eval dataset: 100 valid + 100 invalid + 20 agentic

Swap the embedder: bge-small (384d) -> bge-base (768d)

Retrieval redesign: per-kind pools, relevance threshold, interleave

Let the planner choose the content space: search_docs gains a kind filter

Retrieval: one distance-qualified reference seat instead of two blind ones

Retrieval: guarantee reference pages a seat in the top-k

Apply runtime column migrations from the app pool, not only from builds

Harden the crawler: retry with backoff, size cap, politeness delay

Cut oversized sections at natural seams; embed only the new parts

Add a retrieval benchmark: recall@k + MRR over the v0 questions

Guard: calibrated threshold (0.35) + fail-open on untranslated questions

Guard: replace the injection classifier with embedding-space membership

Guard: multilingual injection classifier first, open model as fallback

App backpressure and cleanups

Share a circuit breaker across LLM calls; cap Retry-After sleeps

Harden the input guard for production

Merge input guard branch (injection + topicality checks)

Serve questions concurrently instead of one-at-a-time (#49)

Add input guard: prompt-injection + topicality check on the user's question

Add a license link under answer citations (#47)

Codebase quality pass: kill silent-truncation hacks, fix latent bugs, cover them (#46)

One dependency file: requirements.txt, read into pyproject dynamically (#45)

Single source of dependency truth: requirements.txt installs the project (#44)

Fix Space runtime crash: requirements.txt missing anthropic + google-genai (#43)

OpenRouter free-tier resilience + GEMINI alias + live HF Space smoke test (#41)

Self-heal LLM: fall back across providers when one is unreachable (#40)