Ummon – Intelligent Codebase Knowledge Graph for AI Assistants

Ummon tackles the core challenge of making large, complex codebases comprehensible to both humans and AI assistants. By automatically generating a semantic knowledge graph that links functions, classes, modules, and even domain concepts, it transforms raw source files into a navigable web of relationships. This representation reduces the cognitive load required to reason about software, enabling developers and AI agents to pinpoint relevant code quickly and accurately.

At its heart, Ummon offers four interlocking capabilities. First, the Knowledge Graph Construction engine parses multiple languages—Rust, Python, JavaScript, and Java—to create a persistent, incremental index of code entities and their interactions. It records calls, imports, dependencies, and file modifications, so the graph stays up‑to‑date without reprocessing unchanged code. Second, the Advanced Querying System lets users retrieve information using either a structured query language or natural‑language prompts. The system can filter, join, and traverse relationships, returning results in text, JSON, CSV, or tree format. Third, the Relevance Agent applies semantic analysis to user descriptions of changes or bugs, ranking files by proximity and graph centrality so that the most pertinent code surfaces first. Finally, Domain Model Extraction leverages large language models to surface business entities and map them onto implementation artifacts, bridging the gap between technical code and domain knowledge.

These features empower a range of real‑world scenarios. In code reviews, an AI assistant can surface all functions related to a security policy by traversing the graph and ranking relevant files. During onboarding, newcomers can query “show me all authentication functions” and receive a concise list of modules to study. When debugging, the relevance agent can surface the most likely culprit files for a reported error, dramatically shortening turnaround time. Moreover, by exposing domain concepts, Ummon allows AI assistants to answer higher‑level questions such as “Which parts of the system handle payment processing?” without needing manual documentation.

Integrating Ummon into AI workflows is straightforward: the server exposes MCP endpoints for indexing, querying, and relevance scoring. A Claude or other LLM client can issue a natural‑language request, have it translated into the structured query language by Ummon’s LLM layer, and receive a ranked list of code snippets or metadata. Because the graph is continuously updated, AI agents can maintain an up‑to‑date mental model of the codebase, leading to more accurate suggestions and fewer misunderstandings.

In summary, Ummon provides a robust, language‑agnostic foundation for AI‑augmented software comprehension. Its combination of incremental graph construction, flexible querying, relevance ranking, and domain extraction gives developers a powerful tool to tame complexity, accelerate development cycles, and enable AI assistants that truly understand the code they help navigate.