Aggressively enforce complete JSON generation in system prompt to prevent truncation

Fix Pydantic validation errors by coercing strings to lists in Protocol parser

Fix Matplotlib backend issue for Strategy Map visualizer

Remove max_tokens limit and strongly discourage long thinking block

Limit K2 reasoning length to prevent JSON truncation

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