What is "cognitive code search" "without embeddings"? Do you mean you accept natural language queries and create embeddings on-the-fly?
edit: That's structural or syntax-aware search.
edit: That's structural or syntax-aware search.
Instead of converting queries to vectors, Mantic.sh uses structural inference, it ranks files based on path components, folder depth, naming patterns, and git metadata.
So "stripe webhook" matches /services/stripe/webhook.handler.ts highly because the path literally contains both terms in logical positions, not because their embeddings are close in vector space.
"Cognitive" just means it mirrors how developers already think about code organization, it encodes intent into paths, so searching paths directly often beats semantic similarity.
This came from a different angle I read about how the human brain operates on ~20 watts yet processes information incredibly efficiently. That got me thinking about how developers naturally encode semantics into folder structures without realizing it.
The "cognitive" framing is because we're already doing the work of organizing code meaningfully Mantic.sh just searches that existing structure instead of recreating it as embeddings. Turns out path-based search is just efficient pattern matching, which explains why it's so fast.
Interesting to hear JetBrains converged on a similar approach from the IDE side though!
This sounds a lot like document search ontop of your specific attributes and you have a custom ranking algorithm.
The 'custom ranking' is inspired by how expert developers navigate codebases: they don't read every file, they infer from structure. /services/stripe/webhook.handler.ts is obviously the webhook handler—no need to embed it.
The innovation is removing unnecessary work (content reading, chunking, embedding) and proving that simple heuristics are faster and more deterministic.
You should benchmark this against other rankers.
The “weights” as you described in Mantic.sh are not neural network weights. They’re deterministic ranking heuristics, similar to what IDE file pickers use. For example, extension weights, path depth penalties, and filename matches. You can see them directly in brain-scorer.ts (EXTENSION_WEIGHTS).
const EXTENSION_WEIGHTS: Record<string, number> = { '.ts': 20, '.tsx': 20, '.js': 15, '.jsx': 15, '.rs': 20, '.go': 20, '.py': 15, '.prisma': 15, '.graphql': 10, '.css': 5, '.json': 5, '.md': 2 };
There’s no LLM involved in the actual search or scoring, it’s a static heuristic engine, not a learned model.
I'd love to have the implementation critiqued on its merits!