Workflow:

Phase 1: Indexing (One-time setup) - indexer.ts

1. Code Discovery 📁

   • walkDirectory() recursively scans your CODE_ROOT directory
   • Filters for .ts, .tsx, .js, .jsx files
   • Skips directories like node_modules, dist, .git

2. Code Chunking ✂️

   • Each file is split into chunks of max 3000 characters
   • chunkContent() splits by lines to keep code contextually intact
   • This prevents hitting OpenAI's token limits

3. Embedding Generation 🧬

   • For each chunk, calls OpenAI's text-embedding-3-small model
   • Converts code text → vector of numbers (embedding)
   • These embeddings capture semantic meaning of the code
   • Rate limited with 100ms delays between API calls

4. Index Storage 💾

   • All chunks saved to code_index.json with structure:

Phase 2: Search (When issue is reported) - search.ts

1. Query Embedding 🔍

   • User's issue description (e.g., "login button not working") → sent to OpenAI
   • Generates embedding vector for the query using same model

2. Similarity Calculation 📊

   • Loads all code chunks from code_index.json
   • Computes cosine similarity between:
     • Query embedding vs. each code chunk embedding
   • Cosine similarity = measures how "similar" two vectors are (0 to 1)

3. Ranking 🥇

   • Sorts all chunks by similarity score (highest first)
   • Returns top 5 most relevant code sections

Phase 3: RCA Generation - engine.ts

1. Prompt Construction 📝

   • Takes user's issue text
   • Adds the top 5 relevant code chunks (with file names & similarity scores)
   • Creates a structured prompt asking GPT-4o-mini to:
     • Analyze the code
     • Find root cause
     • Suggest fixes

2. LLM Call 🤖

3. Response Processing ✅

   • LLM reads the issue + relevant code chunks
   • Returns JSON with:
     • summary: Brief overview
     • thinking: Analysis process
     • root_cause: What's wrong
     • files_mentioned: Which files are involved
     • suggested_fix: Code patch or fix instructions
     • confidence: 0-100 score

QQ: Do LLM reads embeddings then ? Answer: no, LLM to read embeddings, it mains purpose it for searching & keeping the context window shorter, were build that prompt to with filtered code & earlier prompt to give to our llm to perform action on it

QQ: what is cosine similairity? A math formula to measure how similar two vectors are (0 = completely different, 1 = identical). Query vector: [0.12, 0.45, -0.33, ...]

Chunk 1 (github.ts): [0.15, 0.43, -0.31, ...] → Similarity: 0.92 ✓ High! Chunk 2 (slack.ts): [0.78, -0.22, 0.55, ...] → Similarity: 0.34 ✗ Low Chunk 3 (github.ts): [0.11, 0.47, -0.35, ...] → Similarity: 0.89 ✓ High!