How it works
BBM-Atlas is deterministic where it can be and explicit about uncertainty where it cannot. Scroll through the pipeline a repository takes, stage by stage.
01 · Parse
01 · Parsing engine
Tree-sitter reads every file into a language-independent model.
Functions, classes, methods and modules become symbols with qualified names and source spans. There is no LLM in this step, so it is fast, cheap and gives the same result every time.
02 · Structural map
Calls, imports and inheritance are resolved across files.
The structural map links each call site to the definition it reaches and each import to the module it names, tags every edge with how certain the resolution is, and writes it all to a graph database.
03 · Memory & reflection
Summaries, facts and lessons accumulate alongside the code.
Repository memory keeps hierarchical summaries of what each part of the codebase does. The reflection engine records which answers helped and turns them into lessons that adjust future ranking.
04 · Retrieval engine
Four sources are fused into one ranked answer.
A question is answered from structure, memory, the graph and semantic vectors at once. The ranking leans on structural truth, and is boosted by what the question is asking for.
05 · Interfaces
The same engine behind MCP, REST, the CLI and the console.
AI assistants call MCP tools, scripts call the REST API or the CLI, and people use the web console. They share one engine, so they always agree.
06 · Compression layer
Compression shrinks what reaches the model, reversibly.
Content-aware compressors cut tokens from JSON, code, diffs, logs and search results; the originals stay retrievable by hash. A proxy applies it to any OpenAI- or Anthropic-compatible client.
Trust
Each relationship in the graph carries a confidence tag, so an assistant (and you) can weigh an answer, not just read it. Queries can ignore anything below the confidence you need.
Mechanical evidence in the source, from deterministic code paths. Nothing a model produces can ever carry this tag.
A likely binding, such as a call matched by name alone, scored on a fixed rubric so you can see how likely.
A binding the resolver could not make, or one a model flagged as uncertain: kept for review, never guessed.
Storage
Every store sits behind the same interface. A fresh checkout runs hermetically in memory; point BBM-Atlas at real backends when you need persistence and scale.
System of record: repositories, files, symbols, memory
The knowledge graph of engineering relationships
Embeddings for semantic search
Cache and cross-process notifications
Secure defaults
Nothing needs to leave your machine: local embedding models and an in-process LLM are optional extras. And the defaults are safe before you configure anything.
A fresh checkout runs entirely in memory, with no services to set up. Add Postgres, Neo4j, Qdrant and Redis when you are ready.