01
Ingest
Clone the repository and parse every file into an AST with tree-sitter, which covers 100+ languages.
Back to Lore
Technical detail
How Lore is built and served
The indexing pipeline, the three-layer knowledge model, and the tool surface agents actually call.
Pipeline
From repository to queryable knowledge
Seven stages. Stages one through five run on first index; six and seven run continuously after that.
02
Build the graph
Extract symbols and relationships into a graph database, with stable node IDs so later runs can be diffed against earlier ones.
03
Summarize hierarchically
Describe functions, then modules, then subsystems, then the whole repository — each level built on the summaries below it.
04
Discover flows
Traverse the graph from entry points to sinks, then name and describe each flow and the use case it serves.
05
Extract conventions
Read the explicit sources — lint configs, ADRs, contribution guides — and infer the rest from patterns repeated across the code.
06
Index incrementally
On each commit, re-process only what changed and everything that depends on it, so the knowledge base never goes stale.
07
Serve
Expose the knowledge base through MCP tools, and render documentation and diagrams as views over the same data.
MCP tools
Narrow tools, not one ask-anything endpoint
Agents pull exactly the context they need, which keeps answers grounded and token budgets sane.
search_semantic- Natural-language search across the generated summaries.
get_symbol- Fetch a symbol with its definition, location, and immediate context.
get_callers- Everything that calls into a given symbol.
get_callees- Everything a given symbol calls out to.
get_dependencies- Module and package dependencies for an area of the code.
list_flows- Every end-to-end flow discovered in the repository.
get_flow- One flow in full: entry point, path, sinks, and the use case it serves.
get_conventions- The rules that apply to a given area, with evidence.
impact_analysis- What breaks, or needs review, if a given symbol changes.
explain- A composed answer drawing on all three layers, with citations.
Design note
Why three layers rather than one index
Each layer answers a class of question the others answer badly.
A pure vector index handles “how does billing work?” and fails on “what calls this function?” — embeddings blur exactly the precision that question needs. A pure code graph is the reverse: it answers the call question exactly and cannot say what the code is for. Neither knows whether a new file matches how this team writes code.
So the structural layer stays exact, the semantic layer carries meaning, and the conventions layer carries the rules with links to the evidence behind them. The query layer picks the right one, or composes across all three and cites what it used.
Want this pointed at your codebase?
Tell us about the repository and what your team keeps re-explaining to each other, and we'll tell you whether Lore is the right shape for it.