skeg-ollama

Ollama-backed embedding + retrieval pipeline using skeg. Two helpers: OllamaEmbedder (thin embeddings wrapper) and OllamaRetriever (end-to-end embed + store + search). A two-files-of-code path to "I have Ollama running locally, give me a vector DB that does not eat my LLM's RAM."

Talks to skeg over RESP3 (skeg-resp3, port 6379). That is where the TurboQuant tiers live: the native binary protocol cannot name them, because its kind byte 3 means PQ.

Install

pip install skeg-ollama

Pulls in skeg (Python client) and ollama automatically. Server + daemon install is separate:

brew tap skegdb/tap && brew install skeg   # skeg server
ollama serve                               # Ollama daemon
ollama pull nomic-embed-text               # embedding model
skeg-resp3 --data-dir ./data               # the KV+vector backend (RESP3)

Usage

from skeg_ollama import OllamaRetriever

with OllamaRetriever(
    skeg_addr=("127.0.0.1", 6379),
    index_name="my-notes",
    ollama_model="nomic-embed-text",
    # kind defaults to "tq2"; see the tier table below
) as r:
    r.add([
        "the sky is blue",
        "grass is green",
        "cats are fluffy",
    ])
    for text, score in r.search("what colour is the sky?", k=2):
        print(f"  {score:.3f}  {text}")

Output (roughly):

  0.812  the sky is blue
  0.124  grass is green

Quantisation tier (kind)

The tier is fixed when the index is created. Changing it means creating a new index and re-indexing.

Persistence

The retriever's vec_id counter is persisted under the skeg KV side at next-id:{index_name}, so restarting the process keeps the numbering monotonic. The VINDEX itself is persisted via skeg's normal disk path (vindex-<name>/ directory).

When to use this vs skeg-llamaindex

• Use skeg-ollama when you want a tiny, transparent retrieval loop with full control: 200 lines of code top to bottom, no framework.
• Use skeg-llamaindex when you want LlamaIndex's full surface: document chunking, metadata filters, query engines, response synthesis.

Test-suite safety

The pytest suite spawns its own skeg-resp3 fixture; set SKEG_RESP3_BIN to the binary, or the tests skip. Tests create VINDEX entries with names like ollama-test, ollama-batch-..., ollama-empty etc., and drop them at the end. The counter key next-id:<index_name> is left as forensic data after each test drop. If you ever override the fixture to hit an external server, expect those names to collide. The fixture is the safe default.

License

Apache-2.0.