Code Intelligence & Code-Graph Indexing for AI Agents

Tools and emerging approaches for code intelligence and code-graph indexing for AI coding agents from June 2025 through early June 2026, spanning local/embedded indexers (CodeGraph/Caveman-style repo maps, tree-sitter, SQLite and embedded graph stores), enterprise-scale code understanding (SCIP, code knowledge graphs, embeddings+retrieval), LSP-to-MCP bridges such as Serena, and the semantic-vs-syntactic-vs-embedding trade-off.

GPT-5.5
tech
frontier
academic
financial
blogs

Synthesised 2026-06-03

Narrative

The 2025 to early-2026 academic frontier is converging on hybrid repository intelligence: deterministic or static-analysis-backed graph layers for symbol, dependency and navigation tasks; embeddings or sparse retrieval for broad matching; and MCP/LSP bridges for exposing those capabilities to agents. The strongest empirical theme is that purely textual grep-and-read workflows are inefficient at scale, while graph-native or graph-plus-embedding systems can cut tokens and tool calls substantially. The caveat is important: many of the newest practitioner tools are still benchmark-light and vendor-driven.

Semantic, syntactic and embedding methods are not substitutes for one another. Semantic systems are best for symbol navigation, call chains, type-aware lookup and change impact when the repository is buildable or language-server support is available. Syntactic Tree-sitter and AST graphs are easier to deploy locally and across many languages; they are robust and cheap, but miss deeper type and build semantics. Embedding retrieval is strongest for broad natural-language matching and fuzzy discovery, but weak on precise cross-file dependency reasoning unless paired with reranking or graph constraints.

The evidence quality is uneven. The best measured local-indexing evidence in this lane comes from Codebase-Memory's reported token and tool-call savings, alongside METR-style benchmark discipline. LSP-to-MCP tooling is moving quickly, but much of the evidence is still documentation, GitHub repositories or blog posts rather than peer-reviewed evaluation. For large repositories, the most resilient pattern appears to combine symbols, dependencies, commit history and embeddings rather than flattening code into text chunks.

Sources

ID	Title	Outlet	Date	Significance
a1	Codebase-Memory: Tree-Sitter-Based Knowledge Graphs for LLM Code Exploration via MCP	arXiv	2026	Persistent Tree-sitter knowledge graph exposed via MCP; parses 66 languages and reports 10x fewer tokens and 2.1x fewer tool calls than a file-exploration agent on 31 repos.
a2	Repository Intelligence Graph: Deterministic Architectural Map for LLM Code Assistants	arXiv	2026	Deterministic, evidence-backed architectural map of buildable components, aggregators, runners, tests, external packages, and package managers with explicit dependency and coverage edges.
a3	On the Challenges and Opportunities of Learned Sparse Retrieval for Code	arXiv	2026	Introduces SPLADE-Code and argues that learned sparse retrieval can be competitive for code; reports sub-millisecond retrieval on 1M passages with little effectiveness loss.
a4	SemanticForge: Repository-Level Code Generation through Semantic Knowledge Graphs and Constraint Satisfaction	arXiv	2025	Combines dual static-dynamic knowledge graphs, neural graph-query generation, SMT-guided beam search, and incremental KG maintenance.
a5	GRACE: Graph-Guided Repository-Aware Code Completion through Hierarchical Code Fusion	arXiv	2025	Builds a multi-level code graph unifying files, ASTs, call graphs, class hierarchies, and data-flow graphs; hybrid retriever plus graph attention reranker.
a6	RepoScope: Leveraging Call Chain-Aware Multi-View Context for Repository-Level Code Generation	arXiv	2025	Static-analysis-only repository structural semantic graph with call-chain prediction and structure-preserving serialization.
a7	RANGER -- Repository-Level Agent for Graph-Enhanced Retrieval	arXiv	2025	Repository knowledge graph augmented with node text and embeddings; uses Cypher for entity queries and MCTS-guided graph exploration for natural-language queries.
a8	Knowledge Graph Based Repository-Level Code Generation	arXiv	2025	Repository graph representation to improve code search and retrieval for repo-level generation; evaluated on EvoCodeBench.
a9	Beyond Function-Level Search: Repository-Aware Dual-Encoder Code Retrieval with Adversarial Verification	arXiv	2025	Defines RepoAlign-Bench for change-request-driven repo retrieval and proposes a dual-tower retriever with adversarial reflection.
a10	Repository-level Code Search with Neural Retrieval Methods	arXiv	2025	Multi-stage retrieval/reranking for repository-level code search using commit histories plus BM25 and CodeBERT reranking.
a11	RepoGraph: Enhancing AI Software Engineering with Repository-level Code Graph	arXiv	2024	Plug-in repository-level code graph that boosts SWE-bench and CrossCodeEval performance across multiple methods.
a12	GraphCoder: Enhancing Repository-Level Code Completion via Code Context Graph-based Retrieval and Language Model	arXiv	2024	Code Context Graph with control/data/control-dependence edges and coarse-to-fine graph retrieval.
a13	How and Why LLMs Use Deprecated APIs in Code	arXiv	2024	Empirical study showing LLMs rely on code search services and can be influenced by retrieval behavior when using deprecated APIs.
a14	Improving Text Embeddings with Large Language Models	arXiv	2024	LLM-assisted embedding training that improves BEIR/MTEB performance; relevant to embedding-based retrieval quality.
a15	Retrieval Augmented Code Generation and Summarization	arXiv	2021	Early retrieval-augmented code generation/summarization framework (REDCODER).
a16	SCIP Code Intelligence Protocol / Sourcegraph SCIP	Sourcegraph documentation / GitHub	2024	Language-agnostic code indexing protocol for go-to-definition, references, and implementations.
a17	Serena	Open-source MCP toolkit / GitHub	2025	MCP-based coding agent toolkit exposing semantic retrieval and symbol-level editing via LSP integration.
a18	multilspy	GitHub	2024	Python LSP client library intended for applications around language servers.
a19	MCP Bridge: A Lightweight, LLM-Agnostic RESTful Proxy for Model Context Protocol Servers	arXiv	2025	Proxy layer for MCP servers that can simplify access patterns and decouple clients from servers.
a20	CodeSift	Practitioner tool/site	2025	MCP tools for code intelligence claiming reduced-token workflows for agents.
a21	GitHub MCP Server	GitHub repository	2025	Official MCP server supporting repository and workflow intelligence across MCP hosts.
a22	HCAST: Human-Calibrated Autonomy Software Tasks	METR / PDF	2025	Autonomy benchmark suite for software, ML engineering, cybersecurity, and research tasks.
a23	METR preliminary evaluations of Claude 3.7, GPT-4.5, o3/o4-mini, and related frontier-model reports	METR evaluation reports	2025	Comparative agent evaluations on HCAST, SWAA, and RE-Bench, with time-horizon estimates and observations on reward hacking / cheating behaviors.
a24	METR Time-Horizon and Frontier-Risk updates	METR blog / analysis	2025	Time-horizon analyses across software and research tasks; updates on frontier model behavior in task suites.
a25	Context Engineering for AI Agents in Open-Source Software	arXiv	2025	Empirical study of AGENTS.md / AI config files across 466 OSS projects; shows no standard structure yet and strong variation in provided context.