opencode-codebase-index

Stop grepping for concepts. Start searching for meaning.

opencode-codebase-index brings semantic understanding to your OpenCode workflow — and now to any MCP-compatible client like Cursor, Claude Code, and Windsurf. Instead of guessing function names or grepping for keywords, ask your codebase questions in plain English.

📌 Quick Navigation

• ⚡ Quick Start
• 🌐 MCP Server (Cursor, Claude Code, Windsurf, etc.)
• 🎯 When to Use What
• 🧰 Tools Available
• 🎮 Slash Commands
• 📚 Knowledge Base
• 🔄 Reranking
• ⚙️ Configuration
• 🤝 Contributing

👋 Choose Your Path

• I want to try it now → go to Quick Start
• I use Cursor/Claude Code/Windsurf → go to MCP Server setup
• I’m comparing tools and workflows → go to When to Use What
• I’m tuning behavior/cost/performance → go to Configuration
• I want to contribute → go to Contributing

🚀 Why Use This?

• 🧠 Semantic Search: Finds "user authentication" logic even if the function is named check_creds.
• ⚡ Blazing Fast Indexing: Powered by a Rust native module using tree-sitter and usearch. Incremental updates take milliseconds.
• 🌿 Branch-Aware: Seamlessly handles git branch switches — reuses embeddings, filters stale results.
• 🔒 Privacy Focused: Your vector index is stored locally in your project.
• 🔌 Model Agnostic: Works out-of-the-box with GitHub Copilot, OpenAI, Gemini, or local Ollama models.
• 🌐 MCP Server: Use with Cursor, Claude Code, Windsurf, or any MCP-compatible client — index once, search from anywhere.

⚡ Quick Start

1. Install the plugin

   npm install opencode-codebase-index
   

2. Add to opencode.json

   {
     "plugin": ["opencode-codebase-index"]
   }
   

3. Index your codebase Run /index or ask the agent to index your codebase. This only needs to be done once — subsequent updates are incremental.

   Recommended check: run /status after the first index so you can confirm the detected provider/model before you start searching.

4. Start Searching Ask:

   "Find the function that handles credit card validation errors"

Provider selection notes

• Default auto-detect order: Ollama → GitHub Copilot → OpenAI → Google

• Ollama is the preferred zero-cost local option and works especially well for large repos:

  ollama pull nomic-embed-text
  

  {
    "embeddingProvider": "ollama"
  }
  

• GitHub Copilot is a good default if OpenCode already has Copilot auth and you prefer hosted embeddings.

• OpenAI is a good hosted option when you want predictable API behavior and standard cloud setup.

• Google is available if you prefer Gemini-hosted embeddings.

• If /status reports provider or compatibility problems, follow that guidance before using /index force.

🌐 MCP Server (Cursor, Claude Code, Windsurf, etc.)

Use the same semantic search from any MCP-compatible client. Index once, search from anywhere.

1. Install dependencies

   npm install opencode-codebase-index @modelcontextprotocol/sdk zod
   

2. Configure your MCP client

   Cursor (.cursor/mcp.json):

   {
     "mcpServers": {
       "codebase-index": {
         "command": "npx",
         "args": ["opencode-codebase-index-mcp", "--project", "/path/to/your/project"]
       }
     }
   }
   

   Claude Code (claude_desktop_config.json):

   {
     "mcpServers": {
       "codebase-index": {
         "command": "npx",
         "args": ["opencode-codebase-index-mcp", "--project", "/path/to/your/project"]
       }
     }
   }
   

3. CLI options

   npx opencode-codebase-index-mcp --project /path/to/repo    # specify project root
   npx opencode-codebase-index-mcp --config /path/to/config   # custom config file
   npx opencode-codebase-index-mcp                            # uses current directory
   

The MCP server exposes all 9 tools (codebase_search, codebase_peek, find_similar, call_graph, index_codebase, index_status, index_health_check, index_metrics, index_logs) and 4 prompts (search, find, index, status).

The MCP dependencies (@modelcontextprotocol/sdk, zod) are optional peer dependencies — they're only needed if you use the MCP server.

🔍 See It In Action

Scenario: You're new to a codebase and need to fix a bug in the payment flow.

Without Plugin (grep):

• grep "payment" . → 500 results (too many)
• grep "card" . → 200 results (mostly UI)
• grep "stripe" . → 50 results (maybe?)

With opencode-codebase-index: You ask: "Where is the payment validation logic?"

Plugin returns:

src/services/billing.ts:45  (Class PaymentValidator)
src/utils/stripe.ts:12      (Function validateCardToken)
src/api/checkout.ts:89      (Route handler for /pay)

🎯 When to Use What

Scenario	Tool	Why
Don't know the function name	codebase_search	Semantic search finds by meaning
Exploring unfamiliar codebase	codebase_search	Discovers related code across files
Just need to find locations	codebase_peek	Returns metadata only, saves ~90% tokens
Need the authoritative definition site	implementation_lookup	Prioritizes real implementation definitions over docs/tests
Understand code flow	call_graph	Find callers/callees of any function
Know exact identifier	grep	Faster, finds all occurrences
Need ALL matches	grep	Semantic returns top N only
Mixed discovery + precision	/find (hybrid)	Best of both worlds

Rule of thumb: codebase_peek to find locations → Read to examine → grep for precision. For symbol-definition questions, use implementation_lookup first.

📊 Token Usage

In our testing across open-source codebases (axios, express), we observed up to 90% reduction in token usage for conceptual queries like "find the error handling middleware".

Why It Saves Tokens

• Without plugin: Agent explores files, reads code, backtracks, explores more
• With plugin: Semantic search returns relevant code immediately → less exploration

Key Takeaways

1. Significant savings possible: Up to 90% reduction in the best cases
2. Results vary: Savings depend on query type, codebase structure, and agent behavior
3. Best for discovery: Conceptual queries benefit most; exact identifier lookups should use grep
4. Complements existing tools: Provides a faster initial signal, doesn't replace grep/explore

When the Plugin Helps Most

• Conceptual queries: "Where is the authentication logic?" (no keywords to grep for)
• Unfamiliar codebases: You don't know what to search for yet
• Large codebases: Semantic search scales better than exhaustive exploration

🛠️ How It Works

graph TD
    subgraph Indexing
    A[Source Code] -->|Tree-sitter| B[Semantic Chunks]
    B -->|Embedding Model| C[Vectors]
    C -->|uSearch| D[(Vector Store)]
    C -->|SQLite| G[(Embeddings DB)]
    B -->|BM25| E[(Inverted Index)]
    B -->|Branch Catalog| G
    end

    subgraph Searching
    Q[User Query] -->|Embedding Model| V[Query Vector]
    V -->|Cosine Similarity| D
    Q -->|BM25| E
    D --> F[Hybrid Fusion RRF/Weighted]
    E --> F
    F --> X[Deterministic Rerank]
    G -->|Branch + Metadata Filters| X
    X --> R[Ranked Results]
    end

1. Parsing: We use tree-sitter to intelligently parse your code into meaningful blocks (functions, classes, interfaces). JSDoc comments and docstrings are automatically included with their associated code.

Supported Languages (Tree-sitter semantic parsing): TypeScript, JavaScript, Python, Rust, Go, Java, C#, Ruby, PHP, Apex, Bash, C, C++, JSON, TOML, YAML

Additional Supported Formats (line-based chunking): TXT, HTML, HTM, Markdown, Shell scripts

Default File Patterns:

**/*.{ts,tsx,js,jsx,mjs,cjs}    **/*.{py,pyi}
**/*.{go,rs,java,kt,scala}      **/*.{c,cpp,cc,h,hpp}
**/*.{rb,php,inc,swift}         **/*.{vue,svelte,astro}
**/*.{sql,graphql,proto}        **/*.{yaml,yml,toml}
**/*.{md,mdx}                   **/*.{sh,bash,zsh}
**/*.{txt,html,htm}              **/*.{cls,trigger}

Use include to replace defaults, or additionalInclude to extend (e.g. "**/*.pdf", "**/*.csv").

Max File Size: Default 1MB (1048576 bytes). Configure via indexing.maxFileSize (bytes). 2. Chunking: Large blocks are split with overlapping windows to preserve context across chunk boundaries. 3. Embedding: These blocks are converted into vector representations using your configured AI provider. 4. Storage: Embeddings are stored in SQLite (deduplicated by content hash) and vectors in usearch with F16 quantization for 50% memory savings. A branch catalog tracks which chunks exist on each branch. 5. Hybrid Search: Combines semantic similarity (vectors) with BM25 keyword matching, fuses (rrf default, weighted fallback), applies deterministic rerank, then filters by current branch/metadata.

Performance characteristics:

• Incremental indexing: ~50ms check time — only re-embeds changed files
• Smart chunking: Understands code structure to keep functions whole, with overlap for context
• Native speed: Core logic written in Rust for maximum performance
• Memory efficient: F16 vector quantization reduces index size by 50%
• Branch-aware: Automatically tracks which chunks exist on each git branch
• Provider validation: Detects embedding provider/model changes and requires rebuild to prevent garbage results

🌿 Branch-Aware Indexing

The plugin automatically detects git branches and optimizes indexing across branch switches.

How It Works

When you switch branches, code changes but embeddings for unchanged content remain the same. The plugin:

1. Stores embeddings by content hash: Embeddings are deduplicated across branches
2. Tracks branch membership: A lightweight catalog tracks which chunks exist on each branch
3. Filters search results: Queries only return results relevant to the current branch

Benefits

Scenario	Without Branch Awareness	With Branch Awareness
Switch to feature branch	Re-index everything	Instant — reuse existing embeddings
Return to main	Re-index everything	Instant — catalog already exists
Search on branch	May return stale results	Only returns current branch's code

Automatic Behavior

• Branch detection: Automatically reads from .git/HEAD
• Re-indexing on switch: Triggers when you switch branches (via file watcher)
• Legacy migration: Automatically migrates old indexes on first run
• Garbage collection: Health check removes orphaned embeddings and chunks

Storage Structure

.opencode/index/
├── codebase.db           # SQLite: embeddings, chunks, branch catalog, symbols, call edges
├── vectors.usearch       # Vector index (uSearch)
├── inverted-index.json   # BM25 keyword index
└── file-hashes.json      # File change detection

File Exclusions

The following files/folders are excluded from indexing by default:

• Hidden files/folders: Files starting with . (e.g., .github, .vscode, .env)
• Build folders: Folders containing "build" in their name (e.g., build, mingwBuildDebug, cmake-build-debug)
• Default excludes: node_modules, dist, vendor, __pycache__, target, coverage, etc.

🧰 Tools Available

The plugin exposes these tools to the OpenCode agent:

codebase_search

The primary tool. Searches code by describing behavior.

• Use for: Discovery, understanding flows, finding logic when you don't know the names.
• Example: "find the middleware that sanitizes input"
• Ranking path: hybrid retrieval → fusion (search.fusionStrategy) → deterministic rerank (search.rerankTopN) → filters

Writing good queries:

✅ Good queries (describe behavior)	❌ Bad queries (too vague)
"function that validates email format"	"email"
"error handling for failed API calls"	"error"
"middleware that checks authentication"	"auth middleware"
"code that calculates shipping costs"	"shipping"
"where user permissions are checked"	"permissions"

codebase_peek

Token-efficient discovery. Returns only metadata (file, line, name, type) without code content.

• Use for: Finding WHERE code is before deciding what to read. Saves ~90% tokens vs codebase_search.
• Ranking path: same hybrid ranking path as codebase_search (metadata-only output)
• Example output:

  [1] function "validatePayment" at src/billing.ts:45-67 (score: 0.92)
  [2] class "PaymentProcessor" at src/processor.ts:12-89 (score: 0.87)
  
  Use Read tool to examine specific files.
  

• Workflow: codebase_peek → find locations → Read specific files

implementation_lookup

Definition-first lookup. Jumps to the authoritative definition site for a symbol or natural-language definition query.

• Use for: "Where is X defined?", symbol-definition requests, and cases where you want the implementation site rather than all usages.
• Behavior: Prefers real implementation files over tests, docs, examples, and fixtures.
• Fallback: If nothing authoritative is found, use codebase_search for broader discovery.

find_similar

Find code similar to a provided snippet.

• Use for: Duplicate detection, refactor prep, pattern mining.
• Ranking path: semantic retrieval only + deterministic rerank (no BM25, no RRF).

index_codebase

Manually trigger indexing.

• Use for: Forcing a re-index or checking stats.
• Parameters: force (rebuild all), estimateOnly (check costs), verbose (show skipped files and parse failures).

index_status

Checks if the index is ready and healthy.

• Recommended workflow: run this after /index to confirm the detected provider/model and whether the index is ready to search.

index_health_check

Maintenance tool to remove stale entries from deleted files and orphaned embeddings/chunks from the database.

index_metrics

Returns collected metrics about indexing and search performance. Requires debug.enabled and debug.metrics to be true.

• Metrics include: Files indexed, chunks created, cache hit rate, search timing breakdown, GC stats, embedding API call stats.

index_logs

Returns recent debug logs with optional filtering.

• Parameters: category (optional: search, embedding, cache, gc, branch), level (optional: error, warn, info, debug), limit (default: 50).

call_graph

Query the call graph to find callers or callees of a function/method. Automatically built during indexing for TypeScript, JavaScript, Python, Go, and Rust.

• Use for: Understanding code flow, tracing dependencies, impact analysis.
• Parameters: name (function name), direction (callers or callees), symbolId (required for callees, returned by previous queries).
• Example: Find who calls validateToken → call_graph(name="validateToken", direction="callers")

add_knowledge_base

Add a folder as a knowledge base to be indexed alongside project code.

• Use for: Indexing external documentation, API references, example programs.
• Parameters: path (folder path, absolute or relative), reindex (optional, default true).
• Example: add_knowledge_base(path="/path/to/docs")

list_knowledge_bases

List all configured knowledge base folders and their status.

remove_knowledge_base

Remove a knowledge base folder from the index.

• Parameters: path (folder path to remove), reindex (optional, default false).
• Example: remove_knowledge_base(path="/path/to/docs")

🎮 Slash Commands

The plugin automatically registers these slash commands:

Command	Description
/search <query>	Pure Semantic Search. Best for "How does X work?"
/find <query>	Hybrid Search. Combines semantic search + grep. Best for "Find usage of X".
/call-graph <query>	Call Graph Trace. Find callers/callees to understand execution flow.
/index	Update Index. Forces a refresh of the codebase index.
/status	Check Status. Shows if indexed, chunk count, and provider info.

📚 Knowledge Base

The plugin can index external documentation alongside your project code. The indexed codebase includes:

• Project Source Code — all code files in the current workspace
• API References — hardware API docs, library documentation
• Usage Guides — tutorials, how-to guides
• Example Programs — code samples, demo projects

Adding Knowledge Base Folders

Use the built-in tools to add documentation folders:

add_knowledge_base(path="/path/to/api-docs")
add_knowledge_base(path="/path/to/examples")

The folder will be indexed into the same database as your project code. All searches automatically include both sources.

Managing Knowledge Bases

list_knowledge_bases          # Show configured knowledge bases
remove_knowledge_base(path="/path/to/api-docs")  # Remove a knowledge base

Configuration Example

Project-level config (.opencode/codebase-index.json):

{
  "knowledgeBases": [
    "/home/user/docs/esp-idf",
    "/home/user/docs/arduino"
  ]
}

Global-level config (~/.config/opencode/codebase-index.json):

{
  "embeddingProvider": "custom",
  "customProvider": {
    "baseUrl": "https://api.siliconflow.cn/v1",
    "model": "BAAI/bge-m3",
    "dimensions": 1024,
    "apiKey": "{env:SILICONFLOW_API_KEY}"
  }
}

Config merging: Global config is the base, project config overrides. Knowledge bases from both levels are merged.

Syncing Changes

• Project code: Auto-synced via file watcher (real-time)
• Knowledge base folders: Manual sync — run /index force after changes

🔄 Reranking

The plugin supports API-based reranking for improved search result quality. Reranking uses a cross-encoder model to rescore the top search results.

Enable Reranking

Add to your config (.opencode/codebase-index.json or global config):

{
  "reranker": {
    "enabled": true,
    "baseUrl": "https://api.siliconflow.cn/v1",
    "model": "BAAI/bge-reranker-v2-m3",
    "apiKey": "{env:SILICONFLOW_API_KEY}",
    "topN": 20
  }
}

Reranker Options

Option	Default	Description
enabled	false	Enable reranking
baseUrl	-	Rerank API endpoint
model	-	Reranking model name
apiKey	-	API key (use {env:VAR} for security)
topN	20	Number of top results to rerank
timeoutMs	30000	Request timeout

How It Works

Query → Embedding Search → BM25 Search → Fusion → Reranking → Results

1. Embedding Search: Semantic similarity via vector search
2. BM25 Search: Keyword matching via inverted index
3. Fusion: Combine semantic + keyword results (RRF or weighted)
4. Reranking: Cross-encoder rescores top N results via API
5. Results: Final ranked results

Supported Reranking APIs

Any OpenAI-compatible reranking endpoint. Examples:

• SiliconFlow: BAAI/bge-reranker-v2-m3
• Cohere: rerank-english-v3.0
• Local models: Any server implementing /v1/rerank format

⚙️ Configuration

Zero-config by default (uses auto mode). Customize in .opencode/codebase-index.json:

Full Configuration Example

{
  // === Embedding Provider ===
  "embeddingProvider": "custom",              // auto | github-copilot | openai | google | ollama | custom
  "scope": "project",                         // project (per-repo) | global (shared)

  // === Custom Embedding API (when embeddingProvider is "custom") ===
  "customProvider": {
    "baseUrl": "https://api.siliconflow.cn/v1",
    "model": "BAAI/bge-m3",
    "dimensions": 1024,
    "apiKey": "{env:SILICONFLOW_API_KEY}",
    "maxTokens": 8192,                        // Max tokens per input text
    "timeoutMs": 30000,                       // Request timeout (ms)
    "concurrency": 3,                         // Max concurrent requests
    "requestIntervalMs": 1000,                // Min delay between requests (ms)
    "maxBatchSize": 64                        // Max inputs per /embeddings request
  },

  // === File Patterns ===
  "include": [                                // Override default include patterns
    "**/*.{ts,js,py,go,rs}"
  ],
  "exclude": [                                // Override default exclude patterns
    "**/node_modules/**"
  ],
  "additionalInclude": [                      // Extend defaults (not replace)
    "**/*.{txt,html,htm}",
    "**/*.pdf"
  ],

  // === Knowledge Bases ===
  "knowledgeBases": [                         // External docs to index alongside code
    "/home/user/docs/esp-idf",
    "/home/user/docs/arduino"
  ],

  // === Indexing ===
  "indexing": {
    "autoIndex": false,                       // Auto-index on plugin load
    "watchFiles": true,                       // Re-index on file changes
    "maxFileSize": 1048576,                   // Max file size in bytes (default: 1MB)
    "maxChunksPerFile": 100,                  // Max chunks per file
    "semanticOnly": false,                    // Only index functions/classes (skip blocks)
    "retries": 3,                             // Embedding API retry attempts
    "retryDelayMs": 1000,                     // Delay between retries (ms)
    "autoGc": true,                           // Auto garbage collection
    "gcIntervalDays": 7,                      // GC interval (days)
    "gcOrphanThreshold": 100,                 // GC trigger threshold
    "requireProjectMarker": true,             // Require .git/package.json to index
    "maxDepth": 5,                            // Max directory depth (-1=unlimited, 0=root only)
    "maxFilesPerDirectory": 100,              // Max files per directory (smallest first)
    "fallbackToTextOnMaxChunks": true         // Fallback to text chunking on maxChunksPerFile
  },

  // === Search ===
  "search": {
    "maxResults": 20,                         // Max results to return
    "minScore": 0.1,                          // Min similarity score (0-1)
    "hybridWeight": 0.5,                      // Keyword (1.0) vs semantic (0.0)
    "fusionStrategy": "rrf",                  // rrf | weighted
    "rrfK": 60,                               // RRF smoothing constant
    "rerankTopN": 20,                         // Deterministic rerank depth
    "contextLines": 0,                        // Extra lines before/after match
    "routingHints": true                      // Runtime nudges for local discovery/definition queries
  },
  "reranker": {
    "enabled": false,
    "provider": "cohere",
    "model": "rerank-v3.5",
    "apiKey": "{env:RERANK_API_KEY}",
    "topN": 15,
    "timeoutMs": 10000
  },
  "debug": {
    "enabled": false,                         // Enable debug logging
    "logLevel": "info",                       // error | warn | info | debug
    "logSearch": true,                        // Log search operations
    "logEmbedding": true,                     // Log embedding API calls
    "logCache": true,                         // Log cache hits/misses
    "logGc": true,                            // Log garbage collection
    "logBranch": true,                        // Log branch detection
    "metrics": false                          // Enable metrics collection
  }
}

String values in codebase-index.json can reference environment variables with {env:VAR_NAME} when the placeholder is the entire string value. Variable names must match [A-Z_][A-Z0-9_]*. This is useful for secrets such as custom provider API keys so they do not need to be committed to the config file.

{
  "embeddingProvider": "custom",
  "customProvider": {
    "baseUrl": "{env:EMBED_BASE_URL}",
    "model": "nomic-embed-text",
    "dimensions": 768,
    "apiKey": "{env:EMBED_API_KEY}"
  }
}

Options Reference

Option	Default	Description
embeddingProvider	"auto"	Which AI to use: auto, github-copilot, openai, google, ollama, custom
scope	"project"	project = index per repo, global = shared index across repos
include	(defaults)	Override the default include patterns (replaces defaults)
exclude	(defaults)	Override the default exclude patterns (replaces defaults)
additionalInclude	[]	Additional file patterns to include (extends defaults, e.g. "**/*.txt", "**/*.html")
knowledgeBases	[]	External directories to index as knowledge bases (absolute or relative paths)
indexing
autoIndex	false	Automatically index on plugin load
watchFiles	true	Re-index when files change
maxFileSize	1048576	Skip files larger than this (bytes). Default: 1MB
maxChunksPerFile	100	Maximum chunks to index per file (controls token costs for large files)
semanticOnly	false	When true, only index semantic nodes (functions, classes) and skip generic blocks
retries	3	Number of retry attempts for failed embedding API calls
retryDelayMs	1000	Delay between retries in milliseconds
autoGc	true	Automatically run garbage collection to remove orphaned embeddings/chunks
gcIntervalDays	7	Run GC on initialization if last GC was more than N days ago
gcOrphanThreshold	100	Run GC after indexing if orphan count exceeds this threshold
requireProjectMarker	true	Require a project marker (.git, package.json, etc.) to enable file watching and auto-indexing. Prevents accidentally indexing large directories like home. Set to false to index any directory.
maxDepth	5	Max directory traversal depth. -1 = unlimited, 0 = only files in root dir, 1 = one level of subdirectories, etc.
maxFilesPerDirectory	100	Max files to index per directory. Always picks the smallest files first.
fallbackToTextOnMaxChunks	true	When a file exceeds maxChunksPerFile, fallback to text-based (line-by-line) chunking instead of skipping the rest of the file.
search
maxResults	20	Maximum results to return
minScore	0.1	Minimum similarity score (0-1). Lower = more results
hybridWeight	0.5	Balance between keyword (1.0) and semantic (0.0) search
fusionStrategy	"rrf"	Hybrid fusion mode: "rrf" (rank-based reciprocal rank fusion) or "weighted" (legacy score blending fallback)
rrfK	60	RRF smoothing constant. Higher values flatten rank impact, lower values prioritize top-ranked candidates more strongly
rerankTopN	20	Deterministic rerank depth cap. Applies lightweight name/path/chunk-type rerank to top-N only
contextLines	0	Extra lines to include before/after each match
routingHints	true	Inject lightweight runtime hints for local conceptual discovery and definition lookups. Set to false to disable plugin-side routing nudges.
reranker		Optional second-stage model reranker for the top candidate pool
enabled	false	Turn external reranking on/off
provider	"custom"	Hosted shortcuts: cohere, jina, or custom
model	—	Reranker model name required when enabled
baseUrl	provider default	Override reranker endpoint base URL. cohere → https://api.cohere.ai/v1, jina → https://api.jina.ai/v1
apiKey	—	API key for hosted reranker providers
topN	15	Number of top candidates to send to the external reranker
timeoutMs	10000	Timeout for external rerank requests
debug
enabled	false	Enable debug logging and metrics collection
logLevel	"info"	Log level: error, warn, info, debug
logSearch	true	Log search operations with timing breakdown
logEmbedding	true	Log embedding API calls (success, error, rate-limit)
logCache	true	Log cache hits and misses
logGc	true	Log garbage collection operations
logBranch	true	Log branch detection and switches
metrics	false	Enable metrics collection (indexing stats, search timing, cache performance)

Retrieval ranking behavior

• codebase_search and codebase_peek use the hybrid path: semantic + keyword retrieval → fusion (fusionStrategy) → deterministic rerank (rerankTopN) → optional external reranker (reranker) → filtering.
• When search.routingHints is enabled (default), the plugin adds tiny per-turn runtime hints for local conceptual discovery and definition queries. Conceptual discovery is nudged toward codebase_peek / codebase_search, while definition questions are nudged toward implementation_lookup. Exact identifier and unrelated operational tasks are left alone.
• find_similar stays semantic-only: semantic retrieval + deterministic rerank only (no keyword retrieval, no RRF).
• For compatibility rollbacks, set search.fusionStrategy to "weighted" to use the legacy weighted fusion path.
• When enabled, the external reranker sees path metadata plus a bounded on-disk code snippet for each candidate so it can distinguish real implementations from docs/tests more reliably.
• Retrieval benchmark artifacts are separated by role:
  • baseline (versioned): benchmarks/baselines/retrieval-baseline.json
  • latest candidate run (generated): benchmark-results/retrieval-candidate.json

📏 Evaluation Harness

This repository includes a first-class eval system for retrieval quality with versioned golden sets, compare mode, parameter sweeps, CI budgets, and run artifacts.

Commands

npm run eval
npm run eval:ci
npm run eval:ci:ollama
npm run eval:compare -- --against benchmarks/baselines/eval-baseline-summary.json

CI usage split:

• npm run eval:smoke: harness smoke check with local mock embeddings (used in main CI)
• npm run eval:ci: real quality gate against baseline/budget (for scheduled/manual quality workflow)

For eval-quality.yml, the default CI path uses GitHub Models with the workflow GITHUB_TOKEN plus models: read, so you do not need a separate OpenAI API key just to run the scheduled gate.

That default GitHub Models path uses benchmarks/budgets/github-models.json, which applies stable absolute thresholds instead of the stricter baseline-regression budget used for explicit external providers.

Optional override secrets for another OpenAI-compatible endpoint:

• EVAL_EMBED_BASE_URL
• EVAL_EMBED_API_KEY
• EVAL_EMBED_MODEL (optional, default text-embedding-3-small)
• EVAL_EMBED_DIMENSIONS (optional, default 1536)

If you override the provider, set both EVAL_EMBED_BASE_URL and EVAL_EMBED_API_KEY. Otherwise the workflow falls back to GitHub Models automatically. Override providers continue to use the baseline-driven budget in benchmarks/budgets/default.json.

No OpenAI API access? Use Ollama quality gate locally:

• Config: .github/eval-ollama-config.json
• Script: npm run eval:ci:ollama

Prerequisites: Ollama installed, ollama serve running on 127.0.0.1:11434, and nomic-embed-text pulled.

Examples:

# Run against small golden set
npm run eval -- --dataset benchmarks/golden/small.json

# Compare against baseline
npm run eval:compare -- --against benchmarks/baselines/eval-baseline-summary.json --dataset benchmarks/golden/medium.json

# Sweep retrieval parameters
npm run eval -- --dataset benchmarks/golden/small.json --sweepFusionStrategy rrf,weighted --sweepHybridWeight 0.3,0.5,0.7 --sweepRrfK 30,60 --sweepRerankTopN 10,20

What it reports

• Hit@1, Hit@3, Hit@5, Hit@10
• MRR@10, nDCG@10
• Latency p50/p95/p99
• Token estimates, embedding call counts, estimated embedding cost
• Failure buckets (wrong-file, wrong-symbol, docs-tests-outranking-source, no-relevant-hit-top-k)

Artifacts

Each run writes:

benchmarks/results/<timestamp>/

• summary.json
• summary.md
• per-query.json
• compare.json (when baseline/sweep used)

Golden sets and budgets

• Golden datasets:
  • benchmarks/golden/small.json
  • benchmarks/golden/medium.json
  • benchmarks/golden/large.json
• CI budgets:
  • benchmarks/budgets/github-models.json for the default GitHub Models workflow path
  • benchmarks/budgets/default.json for explicit external provider overrides with baseline comparison

Full docs: docs/evaluation.md

Cross-repo benchmark results snapshot

Recent representative runs (plugin vs ripgrep vs ast-grep) on two medium repos:

Methodology for the snapshot below:

• Dataset: auto-generated cross-repo golden sets for axios + express
• Repeats: 20 per mode
• Aggregation: median metric per tool (then averaged across repos)
• Reindex behavior: when enabled, index reset applies on repeat #1 only; subsequent repeats measure warm-index query behavior
• Sampling note: repository parsing can be capped; benchmark reports include truncation metadata
• ast-grep scope note: sg metrics are computed on its compatible query subset (definition, keyword-heavy) with scoped denominators shown in run reports

Without reindex (--no-reindex, default)

Metric	Plugin	ripgrep	ast-grep (5/10 queries)
Hit@5	50%	5%	100%
MRR@10	0.48	0.04	0.90
nDCG@10	0.48	0.08	0.93
Latency p50 (ms)	17.5	36.9	66.6
Latency p95 (ms)	30.9	44.1	70.7

With reindex (--reindex)

Metric	Plugin	ripgrep	ast-grep (5/10 queries)
Hit@5	50%	5%	100%
MRR@10	0.48	0.04	0.98
nDCG@10	0.48	0.07	0.98
Latency p50 (ms)	17.1	35.9	69.1
Latency p95 (ms)	30.4	43.7	75.1

ast-grep metrics are computed on its compatible query subset only (definition + keyword-heavy, 5/10 queries per repo). Plugin and ripgrep are scored on all 10 queries.

Interpretation:

• ast-grep dominates on its scoped subset (structural definition queries), but only handles 50% of query types. Plugin handles all query types including natural language.
• Plugin leads on rank-sensitive quality (MRR/nDCG) vs ripgrep across all query types.
• ripgrep remains a useful speed-oriented lexical baseline but has significantly lower retrieval relevance for intent-style queries.
• Plugin is the fastest tool at p50 (~17ms), ahead of ripgrep (~36ms) and ast-grep (~67ms).
• Reported numbers are rounded to avoid false precision; use report artifacts for full per-repeat audit trails.

For reproducible setup and commands (including with/without reindex), see:

• docs/benchmarking-cross-repo.md

Embedding Providers

The plugin automatically detects available credentials in this order:

1. GitHub Copilot (Free if you have it)
2. OpenAI (Standard Embeddings)
3. Google (Gemini Embeddings)
4. Ollama (Local/Private - requires nomic-embed-text)

You can also use Custom to connect any OpenAI-compatible embedding endpoint (llama.cpp, vLLM, text-embeddings-inference, LiteLLM, etc.).

Rate Limits by Provider

Each provider has different rate limits. The plugin automatically adjusts concurrency and delays:

Provider	Concurrency	Delay	Best For
GitHub Copilot	1	4s	Small codebases (<1k files)
OpenAI	3	500ms	Medium codebases
Google	5	200ms	Medium-large codebases
Ollama	5	None	Large codebases (10k+ files)
Custom	3	1s	Any OpenAI-compatible endpoint

For large codebases, use Ollama locally to avoid rate limits:

# Install the embedding model
ollama pull nomic-embed-text

// .opencode/codebase-index.json
{
  "embeddingProvider": "ollama"
}

The built-in ollama provider uses Ollama's native /api/embeddings endpoint and is the simplest setup when you want to use nomic-embed-text.

For the built-in Ollama path, the plugin budgets nomic-embed-text against an observed effective input limit of about 2048 tokens, not the model's higher advertised theoretical context. This keeps batching and chunk text generation aligned with real Ollama embedding runtime behavior.

If you want to use a different Ollama embedding model through its OpenAI-compatible API, use the custom provider instead and set customProvider.baseUrl to http://127.0.0.1:11434/v1 so the plugin calls .../v1/embeddings.

📈 Performance

The plugin is built for speed with a Rust native module (tree-sitter, usearch, SQLite). In practice, indexing and retrieval remain fast enough for interactive use on medium/large repositories.

• Typical query latency: ~800-1000ms (mostly embedding API time)
• Incremental indexing: only changed files are re-embedded
• Batch DB operations: significant write-speed improvements for large indexes

For reproducible measurements on your machine, run: npx tsx benchmarks/run.ts.

🎯 Choosing a Provider

Quick recommendation:

• Want local + private + fast indexing → use Ollama
• Already have Copilot and a smaller repo → use GitHub Copilot
• General cloud setup → use OpenAI or Google
• Custom/OpenAI-compatible endpoint → use custom provider

Provider Comparison

Provider	Speed	Cost	Privacy	Best For
Ollama	Fastest	Free	Full	Large codebases, privacy-sensitive
GitHub Copilot	Slow (rate limited)	Free*	Cloud	Small codebases, existing subscribers
OpenAI	Medium	~$0.0001/1K tokens	Cloud	General use
Google	Fast	Free tier available	Cloud	Medium-large codebases
Custom	Varies	Varies	Varies	Self-hosted or third-party endpoints

*Requires active Copilot subscription

Setup by Provider

Set the provider in .opencode/codebase-index.json:

{ "embeddingProvider": "ollama" }

Credentials (if required) are read from environment variables (for example OPENAI_API_KEY or GOOGLE_API_KEY).

Custom (OpenAI-compatible) Works with any server that implements the OpenAI /v1/embeddings API format (llama.cpp, vLLM, text-embeddings-inference, LiteLLM, etc.).

{
  "embeddingProvider": "custom",
  "customProvider": {
    "baseUrl": "{env:EMBED_BASE_URL}",
    "model": "nomic-embed-text",
    "dimensions": 768,
    "apiKey": "{env:EMBED_API_KEY}",
    "maxTokens": 8192,
    "timeoutMs": 30000,
    "maxBatchSize": 64
  }
}

Required fields: baseUrl, model, dimensions (positive integer). Optional: apiKey, maxTokens, timeoutMs (default: 30000), maxBatchSize (or max_batch_size) to cap inputs per /embeddings request for servers like text-embeddings-inference. {env:VAR_NAME} placeholders are resolved before config validation for fields that are actually used and throw if the referenced environment variable is missing or malformed.

Custom Ollama models via OpenAI-compatible API If you are running Ollama locally and want to use an embedding model other than the built-in ollama setup, point the custom provider at Ollama's OpenAI-compatible base URL with the /v1 suffix:

{
  "embeddingProvider": "custom",
  "customProvider": {
    "baseUrl": "http://127.0.0.1:11434/v1",
    "model": "qwen3-embedding:0.6b",
    "dimensions": 1024,
    "apiKey": "ollama"
  }
}

Notes:

• The plugin appends /embeddings, so baseUrl should be http://127.0.0.1:11434/v1, not just http://127.0.0.1:11434.
• Ollama ignores the API key, but some OpenAI-compatible clients expect one, so a placeholder like "ollama" is fine.
• Make sure dimensions matches the actual output size of the model you pulled locally.

⚠️ Tradeoffs

Be aware of these characteristics:

Aspect	Reality
Search latency	~800-1000ms per query (embedding API call)
First index	Takes time depending on codebase size (e.g., ~30s for 500 chunks)
Requires API	Needs an embedding provider (Copilot, OpenAI, Google, or local Ollama)
Token costs	Uses embedding tokens (free with Copilot, minimal with others)
Best for	Discovery and exploration, not exhaustive matching

💻 Local Development

1. Build:

   npm run build
   

2. Register in Test Project (use file:// URL in opencode.json):

   {
     "plugin": [
       "file:///path/to/opencode-codebase-index"
     ]
   }
   

   This loads directly from your source directory, so changes take effect after rebuilding.

🤝 Contributing

For contribution workflow, standards, and release-label requirements, see CONTRIBUTING.md.

If you want to add support for a new language, see docs/adding-language-support.md for the full Rust + TypeScript checklist.

Quick path:

1. Fork + branch
2. Implement + tests
3. Run checks: npm run build && npm run typecheck && npm run lint && npm run test:run
4. Open PR with a release category label

Release process (structured + complete notes)

To ensure release notes reflect all merged work, this repo uses a draft-release workflow.

1. Label every PR with at least one semantic label:
   • feature, bug, performance, documentation, dependencies, refactor, test, chore
   • and (when relevant) semver:major, semver:minor, or semver:patch
   • PRs are validated by CI (Release Label Check) and fail if no release category label is present
2. Let Release Drafter build the draft notes automatically from merged PRs on main.
3. Before publishing:
   • copy/finalize relevant highlights into CHANGELOG.md
   • bump package.json version
   • run: npm run build && npm run typecheck && npm run lint && npm run test:run
4. Publish release from the draft (or via gh release create after reviewing draft content).

PRs labeled skip-changelog are intentionally excluded from release notes.

Project Structure

├── src/
│   ├── index.ts              # Plugin entry point
│   ├── mcp-server.ts         # MCP server (Cursor, Claude Code, Windsurf)
│   ├── cli.ts                # CLI entry for MCP stdio transport
│   ├── config/               # Configuration schema
│   ├── embeddings/           # Provider detection and API calls
│   ├── indexer/              # Core indexing logic + inverted index
│   ├── git/                  # Git utilities (branch detection)
│   ├── tools/                # OpenCode tool definitions
│   ├── utils/                # File collection, cost estimation
│   ├── native/               # Rust native module wrapper
│   └── watcher/              # File/git change watcher
├── native/
│   └── src/                  # Rust: tree-sitter, usearch, xxhash, SQLite
├── tests/                    # Unit tests (vitest)
├── commands/                 # Slash command definitions
├── skill/                    # Agent skill guidance
└── .github/workflows/        # CI/CD (test, build, publish)

Native Module

The Rust native module handles performance-critical operations:

• tree-sitter: Language-aware code parsing with JSDoc/docstring extraction
• usearch: High-performance vector similarity search with F16 quantization
• SQLite: Persistent storage for embeddings, chunks, branch catalog, symbols, and call edges
• BM25 inverted index: Fast keyword search for hybrid retrieval
• Call graph extraction: Tree-sitter query-based extraction of function calls, method calls, constructors, and imports (TypeScript/JavaScript, Python, Go, Rust)
• xxhash: Fast content hashing for change detection

Rebuild with: npm run build:native (requires Rust toolchain)

Platform Support

Pre-built native binaries are published for:

Platform	Architecture	SIMD Acceleration
macOS	x86_64	✅ simsimd
macOS	ARM64 (Apple Silicon)	✅ simsimd
Linux	x86_64 (GNU)	✅ simsimd
Linux	ARM64 (GNU)	✅ simsimd
Windows	x86_64 (MSVC)	❌ scalar fallback

Windows builds use scalar distance functions instead of SIMD — functionally identical, marginally slower for very large indexes. This is due to MSVC lacking support for certain AVX-512 intrinsics used by simsimd.

License

MIT