Update artifacts/api-server/src/routes/model.ts

Update artifacts/api-server/src/routes/model.ts

Update artifacts/api-server/src/routes/model.ts

Update artifacts/api-server/src/routes/model.ts

Update attached_assets/io_model_server.py

Update attached_assets/io_model_server.py

Update attached_assets/io_model_server.py

Update attached_assets/io_model_server.py

Update attached_assets/io_model_server.py

Add streaming model refinement to search results

Update attached_assets/io_model_server.py

Update artifacts/api-server/src/routes/model.ts

Update artifacts/api-server/src/routes/search.ts

Update attached_assets/io_model_server.py

feat: model answers from own knowledge first, web results expand after (parallel execution)

perf+fix: drop runPythonScore call (saves 5-12s), keep source-weight+relevance scoring, penalise news language for general answers

feat: Jaccard semantic deduplication — filters near-duplicate facts across Wikipedia/DuckDuckGo/Wikidata

fix: strong sentence quality filter (citation artifacts, fragments, meta-language) + combined perplexity+relevance ranking

fix: modelDriveAnswer uses distilgpt2 only for scoring, returns real web sentences — no hallucination

fix: QA prompt format (Context/Q/A) so distilgpt2 answers instead of hallucinating

fix: modelDriveAnswer no longer times out — nSamples 1, maxTokens 50/40, pool 20, coherence 0.05

feat: migrate HST Crystalline v8 search engine to io (200+ sources, distilgpt2 refinement)

feat: replace 6.2GB model with distilgpt2 (82M, INT8, instant CPU inference)

feat: add HST knowledge distillation script (never overwrites original 6.2GB checkpoint)

perf: INT8 dynamic quantization (5.8GB→~1.5GB, ~2-3x faster CPU inference) + diag logging

diag: log both decode methods side-by-side to find root cause of garbled output

fix: pass attention_mask to model.generate() — matches Colab **inp exactly

fix: decode full sequence like Colab to eliminate BPE boundary garbage

fix: match Colab Cell 10 params + extend CPU timeouts to 8 min

cleanup: strip dead code (web-scraping, fine-tuning, scoring)

Update attached_assets/io_model_server.py

Update attached_assets/io_model_server.py

Update attached_assets/io_model_server.py

Update attached_assets/io_model_server.py

Update attached_assets/io_model_server.py

Update attached_assets/io_model_server.py

perf: greedy decoding (temp=0.1), 60 tokens — fit within CPU budget, no more timeouts

fix: place io_logo.png in attached_assets/ where @assets alias resolves

feat: AI-only mode — remove all web search, model answers every query directly

feat: replace logo with new io_logo.png

perf: greedy decoding + inference_mode for coherent fast answers on CIF path

fix: ThreadingHTTPServer so health checks never block during inference + all CPU threads for generation

fix: load model in fp16 + free checkpoint dict immediately to stay under 16GB RAM limit

fix: increase inference timeout 30s->180s, reduce nSamples 3->1 and maxTokens 400->100 for CPU feasibility

fix: pass attention_mask explicitly to suppress generation warning

fix: auto-download correct 6.2GB checkpoint if cached file is stale/wrong size

fix: complete model server HTTP handler, remove duplicate daemon, extend health timeout to 900s

Update artifacts/api-server/src/index.ts

Update artifacts/api-server/src/index.ts

Update attached_assets/io_model_server.py