Fix K2 Think Engine JSON parsing schema by enforcing strict JSON template

fix: commando-style prompt to force JSON and stop model self-narration

fix: ultra-robust non-greedy JSON extraction and markdown cleanup

fix: production-ready settings (temp=0, timeout=300) and simplified expert prompt for K2 engine

perf: optimize token usage and context window for K2 engine (10k context, 8k response)

debug: add deep logs and index-based extraction fallback for K2 engine

fix: diamond-shield JSON extraction and repair logic with [RESULT] tag support

chore: major reliability overhaul - added result_data storage, auto-migrations, and fixed JSON parsing

fix: use standard json module for parsing to avoid langchain version conflicts

fix: robust JSON extraction using indices and parse_json_markdown

fix: align ReasoningTrace fields with database schema

fix: bump memory collection version and align with fastembed named vectors

fix: align qdrant vector naming with fastembed default

fix: reliable ArXiv ID lookup and uniformized paper ID prefixes

fix: robust qdrant indexing and K2 think JSON extraction logic

fix: validate ScientificDocument schema with authors and document_type

fix: resolve UnboundLocalError for logger in AnalysisService

fix: add auto-migration for remote_id and summary columns in main.py lifespan

fix: automatic paper persistence, qdrant index fix, and JSON output hardening for K2 Think

fix: use qdrant-client[fastembed] extra for proper integration

fix: remove unused unstructured dependency to resolve onnxruntime version conflicts

fix: upgrade unstructured to resolve onnxruntime conflict with fastembed

fix: resolve pillow dependency conflict with fastembed

refactor: restore LangChain orchestration for robust scientific analysis and chat

feat: replace OpenAI embeddings with local FastEmbed models for semantic memory

docs: finalize architecture documentation and ensure exclusive K2 Think model integration

fix: resolve InFailedSqlTransaction by adding db.rollback() in background task and adding detailed error logging

fix: correct internal schema import in AnalysisService background task

fix: correct Response and FileResponse import location

fix: actually start background analysis task and implement process_analysis in AnalysisService

fix: add 'model' field to AnalysisRequest schema to resolve AttributeError

fix: add missing memory fields to API AnalysisRequest schema to prevent Pydantic ValueError

fix: resolve NameError for 'logger' in analysis routes

feat: include direct pdf_url in search results for pre-analysis download

feat: complete memory system with Research Profile (Level 3)

feat: implement persistent chat memory (Conversation Level 1)

docs: update README with OpenAlex, Semantic Memory, and History features

feat: implement Long-Term Semantic Memory using Qdrant (Approach B)

fix: address user feedback - persistent history, session duration (24h), and auto-save papers to default project

feat: complete OpenAlex integration as 4th discovery source (search + download + discipline mapping)

feat: add OpenAlex service for academic paper discovery

fix: resolve InFailedSqlTransaction aborting auto-migration by using information_schema and AUTOCOMMIT

fix: resolve PydanticSchemaGenerationError caused by invalid datetime type in UserResponse

chore: trigger hugging face rebuild

feat: added manual migration endpoint to force schema update on Supabase

fix: auto-migration at startup to add credits/last_refill_date columns and backfill 2000 credits for existing users

fix: expose credits in UserResponse and sync schema updates

fix: expose credits in UserResponse schema for API sync

feat: backend credit management (2000 daily tokens) with auto-refill logic and deduction API

Add DOAJ and PubMed integrations with Discipline Hub classification engine