Download Latest Version v2.1.23 source code.zip (24.4 MB) Google Add to Preferred Sources
Home / v2.1.19
Name Modified Size InfoDownloads / Week
Parent folder
README.md 2026-09-16 3.1 kB
0
v2.1.19 source code.tar.gz 2026-09-16 24.3 MB
0
v2.1.19 source code.zip 2026-09-16 24.4 MB
0
Totals: 3 Items   48.7 MB 0
  • refactored knowledge graph helpers and api to operate on in-memory kg_data dictionaries instead of a sqlalchemy engine, removing per-call database loads/saves and scope normalization from kg_search_facts, kg_get_stats, kg_evolve_knowledge, kg_link_search, kg_embedding_search, kg_backfill_from_memories, kg_explore_concept, and kg_hybrid_search (acec3f3, 5ce6e13)- kg_evolve_knowledge now updates the provided kg_data dict in place rather than persisting to the database (5ce6e13)
  • kg_link_search and kg_explore_concept now traverse local fact_to_concept_links, concept_links, and fact_to_fact_links structures instead of querying sql kg_links tables (5ce6e13)
  • kg_backfill_from_memories now returns an empty stats dict and no longer reads the memory_lifecycle table or evolves the graph (5ce6e13)
  • kg_hybrid_search rewritten to merge kg_link_search and kg_embedding_search results into a single ranked list, dropping the previous mode/keyword logic (5ce6e13)
  • broadened the get_facts extraction prompt to accept concrete declarative factual statements across any domain and refreshed the examples (acec3f3)
  • added pydantic BaseModel response_format support for ollama, including json schema injection and json output parsing (acec3f3)
  • added orcarouter provider support in the litellm response path with correct model and provider routing (acec3f3)
  • npc_array consensus reduction now auto-selects the model from the first npc spec when omitted and passes the provider through to get_llm_response (5ce6e13)
  • npc_compiler now truncates displayed test-case arguments at 200 characters and results at 400 characters (5ce6e13)
  • fixed internal agent logging calls in npc_compiler.py to use log instead of _log, resolving a NameError on every tool-calling loop iteration (e2873ea)
  • added the natural skill and jinx, which render a jinja prompt template with context and send it to the active npc's llm, storing the response in context['output'] (14c2bbf)
  • updated serve.py /stream endpoint to accept api_url and api_key from request payloads, falling back to npc values, and propagate them to llm and agent calls (e2873ea)
  • renamed optional dependency extra from yap to tts in setup.py, the readme, and installation docs, and bumped package version to 2.1.19 (d1bbed1, acec3f3)
  • expanded .gitignore to cover model weights, checkpoints, media, build outputs, and project metadata files (acec3f3)
  • updated readme code examples: added validation assertions to the knowledge-graph sleep/dream example, corrected Team and Jinx paths to examples/npc_team, and updated SememolutionPopulation to require explicit model and provider arguments (acec3f3)
  • updated readme ecosystem links and research paper titles (acec3f3)- added example jinx file examples/npc_team/jinxes/analyze.jinx (acec3f3)
  • added tests/test_readme_examples.py to extract and execute python code blocks from README.md as pytest tests, with environment variables to skip long and media examples (e2873ea)
Source: README.md, updated 2026-09-16