cartog

name: cartog description: >- Code graph navigation, semantic code search, and impact analysis. Use when the user asks "where is X defined?", "what calls X?", "who imports X?", "what depends on X?", "how is X used?", "where is X called from?", "what breaks if I change X?", "is it safe to change/delete X?", "help me refactor X", "show me the call graph", "trace the execution flow", "find all usages of X", "find code related to...", "show me how X works", "show me the X implementation", "show file structure", "list all functions/classes in a file", "show the inheritance tree", "show class hierarchy", or needs to navigate code, locate definitions, search code by concept or behavior, trace dependencies, assess blast radius of changes, explore how a feature is implemented, support refactoring (rename, extract, move, delete), or explore an unfamiliar codebase. Supports Python, TypeScript/JavaScript, Rust, Go, Ruby, Java, PHP, Dart, Swift, Kotlin, Vue, Svelte, Astro, and Markdown documents, plus React/JSX and Vue/Svelte/Astro component-usage edges.

cartog — Code Graph Navigation Skill

Quick Start

Three states a repo can be in. Detect which one the user is in, then act.

Fresh-repo handling (the most common pitfall)

When .cartog.toml is missing on a git repo, the agent must:

1. ASK the user before running cartog init. Do not run it automatically — it writes a .cartog.toml file in the user's repo.
2. On user YES: run cartog init via Bash, check the exit code, then run cartog index .. Both commands are safe to chain in a single session.
3. On user NO (or "skip"): run cartog index . directly. Cartog will use default config and the index lands at <git-root>/.cartog/db.sqlite. The user can run cartog init later if they want a customized config.

If cartog init returns non-zero (rare — usually a filesystem permission issue), surface the error to the user and do not proceed to cartog index.

Non-interactive sessions (CI, piped): cartog's SessionStart hook exits silently when no .cartog.toml is present. Setting CARTOG_AUTO_INIT=1 bypasses the gate and indexes with defaults.

Running cartog commands while MCP is alive

Both cartog init and cartog index are safe to run via Bash during an active MCP session, but for different reasons:

• cartog init is config-only — it writes .cartog.toml, never touches the database. No writer contention possible.
• cartog index . writes to the same SQLite file the MCP server has open. SQLite WAL mode lets it queue behind any active writer, and cartog's migration writes use a busy-retry loop on SQLITE_BUSY — so concurrent index + serve is safe.
• After indexing, MCP tools pick up the new symbols on the next call (no server restart needed).

If MCP runs with --watch, the watcher will also re-index on file changes (and re-embed when the repo already has embeddings). A manual cartog index . is still safe; it just shares the write-queue.

When two cartog serve instances run against the same DB (e.g. two Claude Code windows on the same project), single-writer election picks one as primary and the others attach read-only. Read-only secondaries refuse cartog_index / cartog_rag_index with a clear message but serve the other 14 MCP tools normally (the 2 write tools are gated; the remaining 14 include cartog_update, which arms a machine-level deferred update rather than a DB write, so it is served even though only 13 tools carry readOnlyHint = true). The secondary auto-promotes to primary within ~10s if the primary process dies.

If cartog_index or cartog_rag_index fails with a read-only error, call cartog_stats and check role (primary vs read-only) and watcher_active (whether the primary is auto-reindexing). That tells you whether to wait for promotion (~10s) or whether the primary's watcher will pick up changes on its own.

After the index is ready:

1. Explore an unfamiliar codebase — cartog map gives a file tree + top symbols ranked by centrality. Start here when onboarding or orienting.
2. Search for anything — cartog rag search "your query" is the default entry point. It handles keywords, natural language, and concept queries in a single call.
3. Recent commits broke something? — cartog changes first, then cartog impact <symbol> on the touched symbols.

When to Use

Use cartog before reaching for grep, cat, or file reads when you need to:

• Orient in a codebase → cartog map [--tokens N] (start here for unfamiliar projects)
• Find code by name, concept, or behavior → cartog rag search "query"
• Search project documentation → cartog rag search "query" --kind document
• Understand the structure of a file → cartog outline <file>
• Find who references a symbol → cartog refs <name> (or --kind calls for just callers)
• See what a function calls → cartog callees <name>
• Assess refactoring impact → cartog impact <name> --depth 3
• Understand class hierarchies → cartog hierarchy <class>
• See file dependencies → cartog deps <file>
• See what changed recently → cartog changes [--commits N]
• Triage a regression ("which of my recent commits broke this?") → cartog changes then cartog impact <symbol> on the symbols listed

How to Run

cartog has two modes — CLI (via Bash) and MCP server. Prefer CLI by default.

Do not fall back to grep/glob if MCP tools are missing — use the Bash tool instead.

Rules for both modes:

• Run independent cartog commands as parallel tool calls — whether MCP or Bash
• Subagents: always use CLI mode via Bash. Skip setup — assume the parent has already indexed

CLI mode: every cartog command in the examples below must be run via a Bash tool call. Do not chain with && or | — use separate Bash calls.

All examples below use CLI syntax. MCP tool names and parameters:

Setup

Before first use, ensure cartog is installed and indexed.

If the project uses a remote embedding provider (.cartog.toml has [embedding] provider = "ollama" or provider = "openai"), that provider supplies the embedding model — but cartog rag setup still downloads the cross-encoder reranker (~150MB default, provider-agnostic; there is no remote reranker). Skip rag setup only if you also disable the reranker ([reranker] provider = "none"); otherwise run it once. The openai provider reads its API key from an env var (api_key_env, default OPENAI_API_KEY), never from .cartog.toml.

The plugin's SessionStart hook handles install + indexing automatically:

• Missing binary: install runs in the background; MCP tools become available on the next session. The user can type /cartog-install to install synchronously, or to retry on failure.
• Binary present, no .cartog.toml: on an interactive session, the hook prints a hint pointing at cartog init. On a non-interactive session it exits silently. See "Fresh-repo handling" above for the agent flow.
• Binary present, .cartog.toml present: the hook runs cartog index . (foreground, typically <1s incremental) then forks cartog rag setup + cartog rag index in the background.

If cartog --version fails (binary missing, broken install, wrong architecture), tell the user to run /cartog-install and explain it installs the cartog binary that matches the plugin version.

If cartog --version shows the installed binary is older than the plugin's pinned version, suggest /cartog-install to upgrade (or cartog self update for users on 0.14.0+). Ask before running either — the user may have pinned an older version deliberately.

Search quality tiers

cartog rag search works at three quality levels depending on setup state:

The setup script runs tier 1 blocking (foreground), then tiers 2 + 3 in the background so Claude is responsive immediately. cartog rag search works at FTS5-only quality the moment the index is ready, and upgrades transparently to tier 2 then tier 3 as the background pipeline completes.

How to tell which tier you're on: inspect the result tags in cartog rag search output. [fts5+vector] means tier 3, [fts5] means tier 1 or 2, rerank=... scores appear from tier 2 onward. See references/query_cookbook.md → "Interpreting results" for the full decoder.

First run: tier 2 downloads 230MB of ONNX models (cached in `/.cache/cartog/models/) in the background. Search keeps working at tier 1 in the meantime; logs go to ~/.cache/cartog/session.log`. Subsequent runs are instant.

Database Location

The index is stored in a SQLite database. cartog resolves the path automatically:

For most projects, no configuration is needed — running cartog index . from any subdirectory will place the DB at the git root automatically.

Non-git projects work too. When git rev-parse finds no repo, cartog falls through to the cwd-based path (priority 4), or honours an explicit --db / CARTOG_DB / .cartog.toml [database] path if set. There's no requirement to be inside a git working tree.

# Override examples
cartog --db /tmp/myproject.db index .
CARTOG_DB=~/.local/share/cartog/proj.db cartog index .

Why cartog Over grep/glob

cartog pre-computes a code graph (symbols + edges) with tree-sitter and stores it in SQLite. Compared to grep/glob:

• Fewer tool calls: 1 command vs 3-6 grep/read cycles
• Transitive analysis: impact --depth 3 traces callers-of-callers — grep can't do this
• Structured results: symbols with types, signatures, and line ranges — not raw text matches

Workflow Rules

1. Before you grep or read a file to understand structure, query cartog first.

2. Search routing — pick the right strategy based on the query:

   A. Semantic search (cartog rag search "<query>") — default for all searches. Returns code only by default; use --kind document for docs or --kind all for both. Handles keyword matching (FTS5), vector similarity, and cross-encoder reranking in a single call. Works for both natural language and keyword-style queries. Always use ONE call with the full query — never split a query into multiple rag search calls.

   cartog rag search "authentication token validation"
   cartog rag search "contract management and timesheet signing"
   cartog rag search "config"
   

   B. Structural search (cartog search <name>) — use only when you need a symbol name to feed into refs, callees, impact, or hierarchy. These commands require exact symbol names, not search results.

   cartog search validate_token
   cartog search AuthService --kind class
   

   Routing rules:

   • Need to find code? → A (rag search) — always
   • Need a symbol name for refs/callees/impact? → B (search) first, then the structural command
   • User already gave an exact symbol name? → call refs/callees/impact directly — skip search

3. When using cartog search to locate a symbol before refs/callees/impact:

   • Exactly one result → use that symbol name and file, proceed.
   • Multiple results, same name, different files → add --file <path> to disambiguate.
   • Multiple results, different names → add --kind <kind> to filter, then re-evaluate.
   • Never pass an ambiguous name to refs/callees/impact — the result will be wrong.

4. Use cartog outline <file> instead of cat <file> when you need structure, not content. Then use Read (with offset/limit) for the specific lines you need — this is more efficient than reading entire files.

5. Before refactoring, run cartog impact <symbol> to see the blast radius.

6. Only fall back to grep/read when cartog doesn't have what you need (e.g., reading actual implementation logic, string literals, config values).

7. After making code changes, run cartog index . to update the graph (LSP auto-detected for accurate edges); add --no-lsp only when you need a faster heuristic-only pass. If MCP is running with --watch, a heuristic re-index already happened automatically — when you need LSP-accurate edges, call the cartog_index MCP tool (its warm servers also catch up the heuristic backlog). In CLI-only sessions run the manual index yourself; while a cartog serve peer is running, cartog index . defers its LSP pass to the peer's warm servers (use cartog index . --force to run LSP locally anyway).

8. If refs/callees/impact look incomplete (fewer results than you expect), re-index with LSP to resolve more edges, then re-run the query. In an MCP session call the cartog_index tool — its warm servers also catch up any deferred backlog. In CLI-only sessions run cartog index . without --no-lsp (add --force if a cartog serve peer is running, since a plain index defers LSP to it).

Do / Don't

DO:

• Use cartog rag search as your default search — it combines FTS5 keyword + vector + reranking in one call
• Use cartog search only to get a symbol name for structural commands (refs, callees, impact, hierarchy)
• Trust that rag search degrades gracefully — FTS5 works even without vector embeddings

DON'T:

• Run cartog search and cartog rag search in parallel for the same query — this wastes a tool call. rag search already includes FTS5 keyword matching internally
• Split one query into multiple rag search calls with rephrased variants — one call is enough. The hybrid search (FTS5 + vector + reranker) handles synonyms and related terms internally
• Block on RAG embedding at setup — background indexing is fine, rag search works immediately with FTS5 + reranker
• Assume rag search requires rag index — it works (at reduced quality) with just cartog index .
• Chain multiple cartog CLI commands with && or | — each invocation opens a fresh SQLite connection with full initialization overhead (PRAGMAs, schema checks, cold cache). Run them as separate tool calls instead
• Pipe cartog output through grep — cartog already returns focused, structured results. Filtering with grep discards context (line numbers, kinds, file paths) and can break && chains when grep finds no match (exit code 1)
• Combine unrelated cartog queries in one bash command — this creates false dependencies and hides failures. See references/query_cookbook.md → "Anti-patterns to avoid" for examples

Commands Reference

Index (build/rebuild)

cartog index .                    # Index current directory (with LSP if available)
cartog index . --no-lsp           # Fast heuristic-only index (~1-4s)
cartog index src/                 # Index specific directory
cartog index . --force            # Re-index all files (ignore cache)

By default, cartog index . auto-detects language servers on PATH and uses them to resolve additional edges. LSP results are persisted in the database — subsequent queries benefit without re-running LSP. Use --no-lsp for fast day-to-day indexing. LSP can be omitted entirely at build time with --no-default-features. When a cartog serve for the same database is running, a plain cartog index . skips its own LSP pass (no cold start) and defers resolution to the serve peer; --force always runs LSP locally.

Search (find symbols by partial name)

cartog search parse                          # prefix + substring match
cartog search parse --kind function          # filter by symbol kind
cartog search config --file src/db.rs        # filter to one file
cartog search parse --limit 10               # cap results

Returns symbols ranked: exact match → prefix → substring. Case-insensitive. Max 100 results.

Valid --kind values: function, class, method, variable, import, interface, enum, type-alias, trait, module, document, plus all (code + docs).

RAG Search (hybrid keyword + semantic)

cartog rag search "authentication token validation"
cartog rag search "error handling" --kind function
cartog rag search "database schema setup" --limit 20
cartog rag search "deployment architecture" --kind document

By default, returns code only. Use --kind document for docs or --kind all for both. Uses hybrid retrieval: FTS5 keyword matching + vector KNN, merged via Reciprocal Rank Fusion. When the cross-encoder model is available, results are re-ranked for better precision.

Outline (file structure)

cartog outline src/auth/tokens.py

Output shows symbols with types, signatures, and line ranges — no need to read the file.

Refs (who references this?)

cartog refs validate_token               # all reference types
cartog refs validate_token --kind calls  # only call sites

Available --kind values: calls, imports, inherits, references, raises, implements, type-of.

Callees (what does this call?)

cartog callees authenticate

Impact (transitive blast radius)

cartog impact SessionManager --depth 3

Shows everything that transitively depends on a symbol up to N hops.

Trace (call path between two symbols)

cartog trace handle_request escalate_to_family   # shortest call path, bodies inline
cartog trace handle_request escalate_to_family --depth 4

Returns the shortest calls path from the first symbol to the second, each hop carrying the caller's body inline. Only statically-resolved calls edges are followed (dynamic dispatch is not traced). Use for "how does A reach B?"; use cartog impact for blast radius and cartog refs for all callers.

Hierarchy (inheritance tree)

cartog hierarchy BaseService
cartog hierarchy BaseService --mermaid   # paste-into-PR diagram

Deps (file imports)

cartog deps src/routes/auth.py
cartog deps src/routes/auth.py --mermaid # graph LR with file as root

Map (codebase overview)

cartog map                               # default 4000 tokens
cartog map --tokens 2000                 # compact
cartog map --tokens 8000                 # detailed
cartog map --mermaid                     # graph TD rooted at "Repo"

File tree + top symbols ranked by reference count (centrality). Use at the start of a session for context loading. --mermaid honors the token budget by stopping before it overflows.

Changes (recently modified symbols)

cartog changes                           # last 5 commits + working tree
cartog changes --commits 10              # last 10 commits
cartog changes --kind function           # only functions

Shows symbols affected by recent git changes, grouped by file.

Context (one-shot task bundle)

cartog context "escalate an alert to family when a senior misses check-in"
cartog context "add OAuth refresh token rotation" --tokens 8000   # default 6000

The primary tool for "how does X work?" and understanding an area: one call returns the most relevant symbols (semantic + keyword search), their 1-hop call neighbors, and high-centrality definitions in the same files — bodies inline, budgeted to fit. Read-equivalent: answer from its bundle instead of a chain of search/refs/outline/Read; drill in with the granular tools only for follow-ups it doesn't cover. Raise --tokens (default 6000) for a whole subsystem. Use cartog search for a single known symbol and cartog trace for a specific call path.

Doctor (environment health check)

cartog doctor                            # check all requirements
cartog --json doctor                     # structured JSON output

Validates git repo, config, database, embedding provider, and reranker. Returns OK / Warn / Error per check and exits with code 1 if any error. Run this when commands fail unexpectedly or after first setup to verify everything is working.

Stats (index summary, savings retention hook)

cartog stats                # files, symbols, edges, languages
cartog stats --savings      # tokens-saved breakdown (see below)
cartog savings              # alias for `cartog stats --savings`

cartog savings is the retention hook. Output shape:

cartog · my-project · 5 queries

████████░░  ~83% tokens saved

Without cartog    ~8.5k tokens   (~1700 / query)
With cartog       ~1.4k tokens   (~280 / query)
──────────────────────────────────────────────
Saved             ~7.1k tokens   (~1420 / query)

By tool (call counts):
     2  search
     1  impact
     1  map
     1  refs

Key facts an agent should know:

• With vs without — Without cartog is an equivalent grep+read flow, baselined at ~1,700 tokens/query. With cartog uses the measured ~280 tokens/query.
• By tool = call counts, not per-tool token savings (the multiplier is flat across all tools). The breakdown shows which navigation patterns the user actually relies on.
• Empty-result queries don't count. A cartog search nonexistent that returns [] is not logged — savings reflect real work, not exploratory pings.
• Only writable databases can log. Secondary read-only MCP attaches skip the write, so multi-MCP setups under-report secondary traffic rather than double-counting it.

Watch (auto re-index on file changes)

cartog watch .                           # watch CWD; auto-embeds if the repo already has embeddings
cartog watch . --rag                     # force re-embed (even on a not-yet-embedded repo)
cartog watch . --debounce 3 --rag-delay 30  # custom timings

Under --watch, embeddings auto-refresh when the repo already has them (run cartog rag index once to opt in) — no --rag needed. Editing a symbol's body invalidates its old embedding so it re-embeds on the next pass. Watch always uses heuristic-only indexing (no LSP) for speed. Previously LSP-resolved edges are preserved in the DB.

Serve (MCP server)

cartog serve                    # MCP server over stdio
cartog serve --watch            # watcher; auto-embeds if the repo has embeddings (all clients get --watch by default)
cartog serve --watch --rag      # force auto-embed even on a not-yet-embedded repo

When an agent calls cartog_index via MCP, LSP servers are started once and kept warm for the session. Subsequent index calls reuse warm servers (~2s instead of a cold 2-15s startup). Background watch re-indexing stays heuristic-only.

Self (upgrade / inspect / rollback / migrate)

cartog self update              # upgrade in place to latest stable
cartog self update --check      # report whether an update exists; exit 1 if outdated
cartog self version             # version + target + install source + last check
cartog self rollback            # restore the previous binary saved at <bin>.old
cartog self migrate-db          # move legacy .cartog.db (+ -wal/-shm/.bak) into .cartog/
cartog self migrate-db --dry-run  # preview the planned moves

User-facing maintenance commands. If the agent observes a "new cartog version available" hint or a stale binary, it can suggest the user run cartog self update. cargo install cartog users get an exit-3 refusal pointing at cargo install cartog --force instead. If the agent sees a one-shot deprecation warning about a legacy .cartog.db, suggest cartog self migrate-db.

Init (scaffold project config — user-facing)

cartog init                     # scaffold .cartog.toml in the current project
cartog init --dry-run           # preview without writing

cartog init is config-only: it writes a commented .cartog.toml template if absent and prints next-steps hints. It does NOT index, does NOT wire MCP, and never overwrites an existing .cartog.toml.

When to suggest it: the user is starting cartog on a fresh repo (no .cartog.toml at the git root). The agent should mention it once, then continue with cartog index to build the graph. CLI-only users stop there; users on Claude Code / Cursor / VS Code / etc. can then run cartog ide.

Wire cartog into editors

Two shapes for the same operation:

# `cartog install` — positional, brew/npm/pip convention. Always non-interactive.
cartog install cursor                         # one editor
cartog install cursor vscode codex            # several editors at once
cartog install                                # all detected editors (no positional = "all")
cartog install cursor --dry-run               # preview without writing
cartog install claude-code --no-watch         # drop --watch from Claude Code args

# `cartog ide` — original, supports the interactive multi-select picker.
cartog ide --yes                              # all detected clients, non-interactive
cartog ide --client cursor --yes              # one client (long-form)
cartog ide --scope project --yes              # only .mcp.json / .cursor/mcp.json / .vscode/mcp.json
cartog ide --dry-run                          # preview without writing

Agent gotcha: cartog ide (bare) runs an interactive multi-select picker. An agent calling it via Bash MUST pass --yes (or --client X / --dry-run), otherwise the command blocks waiting for input. cartog install is always non-interactive, so it's the safe default for agents.

Supported clients: claude-code, claude-desktop, codex, cursor, gemini, opencode, vscode, windsurf, zed. User-scope clients whose config dir is missing are reported as "not installed" and skipped. Existing MCP entries for other servers are preserved (idempotent merge).

When to suggest it: the user explicitly asks to wire cartog into an editor, or mentions one of the supported clients. Do not run it speculatively; the answer "yes" should come from the user before the agent invokes this command.

Config (print resolved configuration)

cartog config                   # human-readable resolved config
cartog config --json            # JSON for parsing

Prints the merged config (defaults + .cartog.toml + env overrides). Useful for the agent to verify which database path, embedding provider, or reranker is active before diagnosing a search-quality issue. Read-only — does not modify anything.

Dev-only utilities

Cartog also ships cartog completions <shell> (shell completion script generator for bash/zsh/fish/powershell/elvish) and cartog manpage (emits a troff-formatted man page on stdout). Agents don't normally need these; surface them only if the user explicitly asks for shell completions or installation of a man page. See docs/usage.md for details.

Token Budget

Use --tokens N to limit output to approximately N tokens (human-readable only, ignored with --json):

cartog --tokens 500 search validate
cartog --tokens 200 outline src/db.rs
cartog --tokens 1000 changes --commits 10

JSON Output

All commands support --json for structured output:

cartog --json refs validate_token
cartog --json outline src/auth/tokens.py
cartog --json rag search "authentication"

Add --compact to strip heavy fields from --json and save tokens: symbol bodies (content), docstrings, and cache hashes are dropped; ids, names, kinds, locations, signatures, and scores are kept (context keeps its budgeted bodies). No-op without --json.

cartog --json --compact rag search "authentication"   # ~60% smaller, still locatable

Edge results (refs, callees, impact, deps, trace) carry a provenance field naming which tier resolved the edge: a heuristic tier (same_file, import_path, same_dir, parent_scope, unique_global, kind_disambig) or an LSP outcome (lsp, lsp_external, lsp_unresolvable). Treat lsp/same_file/import_path as high-confidence; unique_global/kind_disambig as best-effort guesses. Omitted for unresolved edges and indexes built before provenance tracking.

Refactoring Workflow

Before changing any symbol (rename, extract, move, delete):

1. cartog changes — quick check: was this symbol touched in recent commits? Pair with cartog impact <symbol> on anything listed to triage regressions before refactoring.
2. cartog search <name> — confirm exact symbol name and file
3. cartog refs <name> — find every usage
4. cartog impact <name> --depth 3 — transitive blast radius
5. cartog hierarchy <name> — if it's a class, check subclasses too
6. Apply changes, then cartog index . to update the graph (add --no-lsp for a faster heuristic-only pass). Skip this step if MCP is running with --watch — the watcher has already re-indexed; only run it in CLI-only sessions or to force LSP edges on demand.
7. Re-run cartog refs <name> to confirm no stale references remain

For the full 3-phase workflow (heuristic → LSP upgrade → verify), see references/query_cookbook.md → "Assess refactoring scope".

Decision Heuristics

Limitations

• Heuristic resolution is name-based (~25% of edges resolved). With LSP enabled, ~42-81% resolved depending on language. Remaining unresolved edges are mostly calls to external libraries.
• Code languages: Python, TypeScript/JavaScript, Rust, Go, Ruby, Java, PHP, Dart, Swift, Kotlin.
• Frontend SFCs: Vue (.vue), Svelte (.svelte), Astro (.astro) — symbols/edges from the <script>/frontmatter block; template markup is not indexed.
• Frameworks: React/JSX and SFC component usage (<Counter/>) becomes a Calls edge (uppercase/dotted tags only; lowercase HTML elements are skipped).
• Documents: Markdown (.md) — indexed by heading sections for semantic search.
• Does not index string literals, comments (except docstrings), or config values.
• Method resolution is name-based without LSP — foo.bar() resolves bar, not Foo.bar specifically. LSP resolves to the exact type when a language server is available.

LSP limitations

• Default feature: shipped by default. Installs with --no-default-features omit LSP entirely (equivalent to --no-lsp at runtime).
• Auto-detected: if language servers are on PATH, they are used automatically during cartog index. Use --no-lsp to skip.
• Startup latency: language servers typically reach ready in 2-15s on cold cache. The default ready-timeout is 20s — override via CARTOG_LSP_READY_TIMEOUT_SECS for very large projects. Reach for --no-lsp when that startup cost isn't worth it (a quick re-index in a tight edit loop); keep LSP on when edge accuracy matters (refactoring, incomplete refs).
• CLI vs MCP: each cartog index . via Bash spawns and kills LSP servers (cold start). Use cartog serve (MCP mode) for sessions with multiple index calls — it keeps servers warm across tool calls.
• Supported servers: rust-analyzer, pyright-langserver, typescript-language-server, gopls, ruby-lsp, solargraph, jdtls, intelephense (phpactor fallback), dart, sourcekit-lsp, kotlin-language-server, vue-language-server, svelteserver, astro-ls. Install hints shown when servers are missing.
• External crate edges stay unresolved: LSP resolves definitions within the project. Calls to std/external crates remain unresolved regardless.

RAG search limitations

• No substring matching: "valid" does NOT match validate_token. FTS5 is token-based. If rag search returns no results for a known symbol name, fall back to cartog search which supports substring matching.
• Graceful degradation: rag search works without rag setup or rag index (FTS5-only). Quality improves with each setup tier (see Search quality tiers above).
• Scores are relative: rrf_score and rerank_score values are only meaningful for ranking within a single query — don't compare scores across different queries.
• Re-ranking latency: cross-encoder scores all candidates in a single batch ONNX call (up to 50 candidates). Expect ~50-300ms total overhead depending on candidate count (the default jina-turbo reranker is ~3x faster than the older bge-reranker-base).
• Auto re-embed: when cartog upgrades its embedding format (e.g., AST-aware chunking), cartog rag index automatically detects the change and re-embeds all symbols. No --force needed.