Roadmap

Planned improvements for CodexA, organized by priority.

High Priority

RAG Pipeline for LLM Commands

Replace the current "dump context → prompt" approach with a proper Retrieval-Augmented Generation pipeline:

• Semantic retrieval with re-ranking
• Token-aware context assembly
• Source citation in responses
• Configurable retrieval strategies

Async Web & WebSocket Streaming

Migrate the web server from synchronous http.server to an async framework:

• Real-time search result streaming via WebSocket
• Non-blocking request handling
• Server-sent events for long-running operations
• Connection pooling for multi-client scenarios

Precise Token Counting

Replace rough len(text) // 4 estimation with model-specific tokenizers:

• tiktoken for OpenAI models
• Model-specific tokenizers for Ollama models
• Accurate context window budgeting
• Token usage reporting

Medium Priority

Field-Scoped Search

Allow narrowing searches by metadata fields:

codex search "auth" --lang python --symbol-type class --file "src/**"

• Language, symbol type, file path filters
• Pre-filter before vector search for efficiency
• Composable filter expressions

Configurable RRF Weights

Make the Reciprocal Rank Fusion weights tunable:

{
  "search": {
    "rrf_k": 60,
    "vector_weight": 0.7,
    "keyword_weight": 0.3
  }
}

Plugin Sandboxing

Isolate plugins to prevent interference:

• Resource limits (memory, CPU time)
• Restricted filesystem access
• Capability-based permissions
• Graceful error containment

LLM Call Retry Logic

Add resilience for LLM provider calls:

• Exponential backoff with jitter
• Provider failover (OpenAI → Ollama)
• Request timeout configuration
• Rate limit awareness

Low Priority

Fine-Tuned Embedding Models

Train custom embedding models on codebases:

• Domain-specific vocabulary handling
• Language-aware fine-tuning
• Benchmark against general-purpose models

Distributed Indexing

Support indexing across multiple machines:

• Sharded FAISS indices
• Distributed embedding computation
• Merged search results

Contributing

Interested in working on any of these? Check the GitHub issues for related discussions.