Spaces:
Running
Running
Commit History
-
Улучшить выбор LLM-кандидатов оптимизатора
Co-authored-by: Cursor <cursoragent@cursor.com>0e32cca
-
feat(optimizer): stage and goal selection controls
Add per-stage enablement and manual/mixed goal overrides in API, backend optimizer pipeline, and UI with project save/load support. Made-with: Cursor55a7978
-
fix(bm25): strict n-gram remove guidance
Use exact token n-gram counts for bm25 remove span selection and local improvement. Made-with: Cursor865a2ea -
feat(optimizer): add diff highlighting
Expose diff_mode + original snapshot in API, compute sentence-level diff on backend, and render/persist diff view in UI. Made-with: Cursor9e6afbe
-
optimizer: per-goal deficit-based iteration and candidate budget
Made-with: Cursor02e950c -
refactor(optimizer): apply unified per-goal iteration budget across all stages
Build per-stage goal lists and iterate each goal with max_iterations attempts before moving on. Total run budget now scales with number of actionable goals, and docs are updated to match the universal rule. Made-with: Cursorc204306 -
fix(optimizer): keep BERT stage from plateau auto-skip
Prevent automatic stage advancement from BERT after 3 no-progress steps. BERT now continues until per-phrase threshold completion or loop budget exhaustion; docs updated. Made-with: Cursor7c286d9 -
fix(optimizer): require bert target for each phrase
Prevent early BERT stage completion when only one phrase reaches the threshold. The stage now completes only when all tracked phrase scores meet bert_stage_target, and docs reflect the new per-phrase rule. Made-with: Cursor6859d9b
-
fix(ui): apply optimized Title from analysis snapshot + resolved_title in metrics
Made-with: Cursor74af0ae -
fix(optimizer): reject title goal in _llm_edit_chunk; doc deploy check for title_rewrite
Made-with: Cursora8c0c0b -
fix(optimizer): title stage rewrites Title field only, return optimized_title
Made-with: Cursor269a8f1
-
fix(optimizer): define _emit before first call (avoid UnboundLocalError)
Made-with: Cursorb42a56b -
Optimizer: SSE log panel + static app.js (SES-safe nv), no global loader; HF_SES doc
Made-with: Cursor6b11536 -
Revert "Optimizer: SSE progress, cancel job with partial result, UI progress card"
This reverts commit b6a4468ec1517d6b3943366319a5ff2181c7ec81.ba71965 -
Optimizer: SSE progress, cancel job with partial result, UI progress card
Made-with: Cursorb6a4468
-
Revert to pre-progress-bar (ad42da6): remove SSE endpoint, optimizer callbacks, progress UI
Made-with: Cursor1982e6f -
Optimizer: SSE progress stream + UI progress bar during LLM run
Made-with: Cursor59a0de2 -
N-gram: rank by Freq/Avg, extra loop budget, overlapping window chunking
Made-with: Cursorad42da6
-
Enforce per-term ngram loop and underrepresented-only targeting.
Limit n-gram stage targets to underrepresented terms, run three attempts per n-gram before advancing, and document the per-term attempt policy in optimizer principles. Made-with: Cursora52efa4 -
Align n-gram selection rules across summary and optimizer.
Apply stage candidate rules for bi/tri-grams and keyword-derived unigrams, add n-gram target rotation to avoid single-term stalls, and sync summary n-gram logic with optimizer tolerance policy. Made-with: Cursor86b2387 -
Implement n-gram tolerance stage rules and optimizer principles doc.
Apply user-defined n-gram gap policy (20% for avg>=4, 50% for avg<4), optimize n-gram stage toward competitor averages with local-safe acceptance, and add a living optimizer principles document. Made-with: Cursord5937ae -
Add configurable BERT stage target and stage visibility in optimizer.
Allow users to set custom BERT threshold for Stage A completion (e.g. accept 0.61), persist it in UI project state, and expose stage transitions in logs for transparent progression to BM25/Semantic/N-gram/Title. Made-with: Cursor84ebb46
-
Implement staged multi-objective optimization flow after BERT.
Add explicit stage pipeline (BERT -> BM25 -> Semantic -> N-gram -> Title), stage progress/plateau transitions, and expose candidate phrase strategy in debug table for clearer diagnostics. Made-with: Cursorf8eb22b -
Add ensemble phrase strategy mode for optimizer.
Keep user-controlled phrase strategy selection, add optional Ensemble cycling across strategies per iteration, and surface strategy usage in debug/output for comparison. Made-with: Cursord2ba52e -
Add configurable phrase strategy mode for LLM optimizer.
Expose Auto/Distributed/Exact phrase strategy in UI and API, then enforce it in prompt generation and response metadata to reduce unnatural exact-phrase stuffing. Made-with: Cursor09a2c0e -
Add multi-objective candidate utility for optimizer decisions.
Introduce a dynamic utility function that balances BERT phrase gains with cross-metric regression penalties, and use it in candidate selection/ranking to preserve future optimization capacity. Update full documentation to describe the utility-driven acceptance logic. Made-with: Cursordd4e1d6 -
Revert "Improve universal chunk selection with structural and diversity signals."
This reverts commit 9e53d78ff6c3ca60597be1793bb58323ecb025cc.57bdda9 -
Improve universal chunk selection with structural and diversity signals.
Add language-agnostic structural noise scoring, entity continuity bonus, novelty penalty for repeatedly edited sentences, and zone-based diversity rotation for span selection. Update full documentation to reflect the new universal ranking strategy. Made-with: Cursor9e53d78 -
Refine LLM prompt for semantic-vector chunk optimization.
Update optimizer prompts to enforce subject preservation and sentence-by-sentence meaning retention for rewrites, allow exact or distributed core-term usage for BERT goals, and collect one-line rationale from the model. Surface rationale in debug UI and synchronize full documentation. Made-with: Cursoraa803bd -
Apply local-first accumulation for chunk-level optimizer progress.
When no promotable candidate is found but guardrail-valid local improvements exist, apply the best local edit immediately and continue to the next chunk. Add explicit applied_local_progress logging and document the accumulation strategy. Made-with: Cursorf65551f
-
Accept incremental BERT phrase gains during optimization.
Treat small positive document-level phrase deltas as progress for BERT goals, prioritize candidates by bert phrase delta in selection, and expose bert delta in optimizer debug tables. Update full documentation for the new threshold and staged acceptance behavior. Made-with: Cursor00148a1 -
Add semantic gap tolerance and detailed gap attribution.
Use a 15% relative tolerance plus minimum absolute delta when counting semantic gaps, expose semantic gap sum/terms in metrics, and include gap-term add/remove deltas for candidate diagnostics. Update full documentation for the new semantic gap parameters. Made-with: Cursore884b6b -
Improve optimizer attribution and multi-edit batch search.
Add per-candidate metric deltas to explain global score drops, strengthen BERT chunk selection toward non-noise prose, keep threshold-based BERT goal gating, and evaluate non-conflicting queued edit combinations of size 2-4 for better aggregate gains. Update debug UI and full documentation accordingly. Made-with: Cursor0cd3c0b -
Add deep LLM-input diagnostics for chunk optimization debugging.
Record prompt payload context per candidate, chunk relevance before/after, term-level diffs, and clearer valid-vs-promotable rejection reasons to analyze why edits fail despite local rewrites. Update debug UI columns and full functional documentation accordingly. Made-with: Cursora05c226 -
Implement two-stage optimizer acceptance with multi-chunk batching.
Add chunk-level relevance delta scoring, multi-span candidate generation per step, queue-and-batch application for locally improved edits, and noise-aware span ranking. Extend optimizer debug UI and full documentation with local chunk metrics and staged acceptance flow. Made-with: Cursor78494fc -
Improve optimizer span selection to avoid repeated chunk edits.
Add goal-aware sentence ranking, span-variant rotation, attempted-span memory, and richer span diagnostics in logs so cascade levels explore different text regions instead of looping on one chunk. Update full documentation with the new cascade selection behavior. Made-with: Cursor0472bc1 -
Improve optimizer candidate validation and BERT goal tracking.
Add cascade-aware sentence limits, pre-validate LLM outputs for quality issues, and evaluate BERT gains by phrase score delta to improve acceptance decisions. Remove obsolete screenshot assets from repository. Made-with: Cursor270367c
-
Implement cascade state-machine for optimizer edit escalation
Add level-based optimization flow (rewrite -> wider rewrite -> insert -> broader rewrite), track per-step baseline and operation metadata, and escalate automatically after repeated failed attempts. Made-with: Cursor513cc7e -
Add detailed optimizer debug trace for candidate-level visibility
Expose per-iteration sentence edits, candidate scores, validity flags, rejection reasons, and chosen variants in UI to diagnose why optimization steps are accepted or rejected. Made-with: Cursor42ad6d7 -
Tune optimizer acceptance policy and add optimization modes
Improve edit acceptance by supporting conservative/balanced/aggressive modes, continuing after rejected steps, prioritizing goal-level gains, and surfacing rejection reasons in iteration logs. Made-with: Cursoreb8813a -
Add LLM optimizer tab with iterative constrained text improvements
Introduce an OpenAI-compatible optimizer workflow with user-provided API key, local sentence-level edits, per-iteration rescoring, and before/after metric tracking in a dedicated UI tab. Made-with: Cursordc860ce