added llama-server inference support

[pvasilek]

There is support for llama-server inferenced embedding models, like
"llama-server.exe" --model %MODEL_PATH% --alias "text-embedder" --embeddings --jinja --port 11435

[mraza007]

1. On src/memory/embeddings/llama.py (return lines in embed() and search()):
   resp.json()[0]["embedding"][0] returns a single float, not the full embedding vector.
   This class advertises list[float], so this should likely return the full embedding array (same for both methods).
2. On src/memory/embeddings/llama_nomic.py (return lines in embed() and search()):
   Same issue here: resp.json()[0]["embedding"][0] returns one dimension only.
   Please return the full vector, otherwise downstream vector search/dimension logic will break.
3. On src/memory/embeddings/base.py (new abstract search()):
   Making search() abstract is a breaking interface change for existing embedding providers/test doubles that only implement embed().
   This currently breaks fixture instantiation (FakeEmbeddingProvider) in tests.
   Suggestion: add a default search() implementation that delegates to embed() for backward compatibility.
4. On tests/test_search.py (assertions around embedding calls):
   Tests still assert embed() is called, but implementation now calls search().
   Please update these assertions (or keep compatibility via default search=embed) so tests reflect the new contract.

[pvasilek]

It is not a single float as llama-server return array of arrays. Check the log below.

srv  log_server_r: done request: POST /embeddings 127.0.0.1 200
srv  log_server_r: request:  {"model":"text-embedder","content":"Switched to JWT auth Replaced session cookies with JWT Needed stateless auth for API All endpoints now require Bearer token auth jwt"}
srv  log_server_r: response: [{"index":0,"embedding":[[-0.0443197600543499,...,-0.05202531814575195]]}]

As I see 'search = embed' exists in the old classes. It does not work as expected?

[mraza007]

Thanks, this helps. You’re right that your llama-server response shape is nested (embedding: [[...]]), so indexing [0] can be expected there

---

Please add tests for the new search() contract before merge.

I’m still seeing two regressions on current PR head:

1. EmbeddingProvider.search() is abstract now, which breaks existing providers/test doubles that only implement embed() (fixture error in tests/conftest.py).
2. tests/test_search.py still expects embed() calls, but implementation now calls search().

Can you update/add tests to cover this explicitly?

[pvasilek]

I am not a python programmer and do not have pytest here. So, for me that is quite complex task, sorry.

[mraza007]

Sounds good
I'll take care of the tests