Skip to content

Hybrid Vector Retrieval & Dense Embeddings

Accurate codebase retrieval requires combining two search modalities: 1. Lexical Search (Exact Identifiers): Matches exact variable names, class names, method symbols, and file paths. 2. Dense Semantic Search (Natural Language): Matches intent, architectural concepts, and natural language descriptions.

TLDRGraph implements a hybrid retrieval engine combining lexical TF-IDF with local FastEmbed ONNX dense embeddings.

The Dual-Channel Retrieval Architecture

graph TD
    Query["User / Agent Query"] --> Lexical["Lexical TF-IDF Engine"]
    Query --> Dense["FastEmbed ONNX (BAAI/bge-small-en-v1.5)"]
    Lexical --> Scoring["Reciprocal Rank Fusion (RRF) & Layer Weighting"]
    Dense --> Scoring
    Scoring --> Slices["Layer-Grounded Execution Slices (~8,000 Tokens)"]

1. Lexical Search (vector_tfidf.py)

  • Tokenizes symbols using camelCase, snake_case, and dotted package conventions.
  • Weighs symbol identifiers, file path components, and docstring terms using BM25/TF-IDF scoring.

2. Dense Embeddings (dense_embedder.py)

  • Employs FastEmbed running the optimized BAAI/bge-small-en-v1.5 ONNX model.
  • Runs entirely locally on CPU: Requires no GPU, no CUDA, and no PyTorch installation.
  • Quantized and blazingly fast: query latency averages under 15ms.

Layer-Grounded Execution Slices

Standard RAG systems blindly return chunked paragraphs from random files.

TLDRGraph leverages its dynamic layer awareness: - When a query retrieves a function in the business layer, it traverses the graph upwards to find the entry point route in the API layer and downwards to find the database model. - It bundles these connected symbols into Layer-Grounded Slices (~8,000 tokens). - This ensures the LLM receives the full context of how data flows across all architectural boundaries.