Micro Brain sounds like a tiny model. V1 is closer to a deterministic semantic retrieval library.
It runs locally, requires no embedding API, and identical source plus compiler settings must reproduce the same logical index.
248 cards become 1,118 retrieval units
The current source contains 248 cards: 225 canonical bridges, 23 open concepts, 25 guarded formulas, 159 relations and 56 confusion pairs.
The compiler splits cards into definition, terminology, positive, negative, relation and formula/evidence units. The current index contains 1,118 units and 2,048 sparse features. The compiled JSON is about 5.63 MB and roughly 1.73 MB gzipped.
Each query is retrieved twice
BM25 preserves exact business terminology. A TF-IDF/LSA projection provides a deterministic dense view for semantic interpolation.
RRF merges ranks, not confidence
BM25 and cosine similarity are not pretending to share one confidence scale. Reciprocal Rank Fusion only decides which concepts should be inspected first.
RRF = 1 / (k0 + sparse_rank)
+ 1 / (k0 + dense_rank)Retrieval similarity remains provenance, never semantic authority.
The query is not allowed to know the answer
File names, corpus IDs, expected answers and hidden customer dictionaries are forbidden from the query signature. Bounded representative values remain evidence witnesses, not full-column truth.
Determinism is part of the product
The compiler records corpus, tokenizer, vector algorithm, dimensions and hashes. Reproducibility matters more here than mysterious “smartness.”
The value of this 5.6 MB brain is not that it can decide everything. It is that it can recall more while still handing authority back to the governed pipeline.
No approved comments yet.