Local semantic code indexer and LLM context packer written in Odin.

OG implements a from-scratch RAG retrieval stack - chunking, indexing, hybrid BM25 + vector search, and token-budgeted context assembly - in a single static binary. No Python, LangChain, Chroma, or vector database required.

Most "AI code search" are thin wrappers around cloud APIs and vector DBs. OG is light, portable and only uses retrieval primitives:

• Code-aware chunking at function/signature boundaries
• Hybrid search - BM25 for symbols + embeddings for intent
• Reciprocal Rank Fusion to merge keyword and semantic rankings
• Context packing with token budgets for agent workflows
• Flat-file index under .og/ - portable, zero daemon

curl -Ls https://raw.githubusercontent.com/pasha1980/og/main/install/install.sh | bash

Prerequisites

• Odin compiler (dev-2026-07 or newer)
• clang (for linking)

git clone https://github.com/pasha1980/og
cd og
make build
# or: odin build ./src -out:og
mv og ~/.local/bin/og

# BM25-only (offline, no external services)
og index /path/to/repo

# With Ollama embeddings (recommended)
ollama pull nomic-embed-text
og index /path/to/repo --embed

# With local GGUF model (no Ollama)
og index /path/to/repo --gguf /path/to/model.gguf

og search "where is JWT validated"
og search --json -k 5 "authentication middleware"

og context "fix refresh token expiry bug" --budget 4000 --format cursor

Copy the markdown output into Cursor, Claude, or any agent.

./og diff-index /path/to/repo

Re-indexes only files changed in git.

og index
  └─ walker      gitignore-aware file traversal
  └─ chunker     language-aware code splitting
  └─ embed       Ollama HTTP / GGUF token embeddings
  └─ index       binary persistence (.og/)

og search / context
  └─ search      BM25 + cosine similarity + RRF merge
  └─ packer      token-budgeted markdown output

.og/
  manifest.json   repo metadata, embed model, dimensions
  chunks.bin      length-prefixed chunks + line metadata
  bm25.idx        inverted index (binary)
  vectors.bin     f32 embeddings (optional)

The GGUF backend loads token_embd weights and produces hash-token mean-pooled embeddings - useful offline and as a learning reference for quantized inference.

On this repository:

og index .
og search "reciprocal rank fusion"
og context "how does gitignore parsing work" --budget 2000

BM25 excels at symbol lookup (ValidateJWT, rrf_merge). With --embed, semantic queries like "where do we handle expired sessions" improve recall.

make test
# or: odin test ./src/tests/ -file

MIT - see LICENSE.