Took me a while to figure why it wasnt working properly with my local ollama

Summary

• Root cause: _call_openai_compat never passes num_ctx to Ollama. Ollama's default is 2048 tokens, so any extraction chunk exceeding that returns content = "", which _parse_llm_json reports as Expecting value: line 1 column 1 (char 0).
• Impact: All Ollama extractions silently produce 0 nodes / 0 edges unless you work around it by shrinking token_budget below Ollama's default context (which defeats the purpose).
• Fix: Two lines in _call_openai_compat — detect backend == "ollama", inject extra_body={"options": {"num_ctx": int(os.environ.get("GRAPHIFY_OLLAMA_NUM_CTX", 32768))}}. The env var already exists in the CLI, just needs to reach the API layer.

Details

The fix is in graphify/llm.py in the _call_openai_compat function. Here's the exact problem and solution:

The bug: When calling Ollama, graphify never passes num_ctx to override Ollama's default context window (2048 tokens). Any chunk larger than ~2048 tokens silently returns an empty response. The GRAPHIFY_OLLAMA_NUM_CTX env var exists in the CLI but is never wired into the Python API.

Current code (around the extra_body section):

# Kimi-k2.6 is a reasoning model — disable thinking so content isn't empty
if "moonshot" in base_url:
    kwargs["extra_body"] = {"thinking": {"type": "disabled"}}
resp = client.chat.completions.create(**kwargs)

Fixed code:

# Kimi-k2.6 is a reasoning model — disable thinking so content isn't empty
if "moonshot" in base_url:
    kwargs["extra_body"] = {"thinking": {"type": "disabled"}}

# Ollama requires explicit num_ctx — default is 2048 which causes silent
# empty responses for any chunk exceeding that. GRAPHIFY_OLLAMA_NUM_CTX
# is already documented for the CLI but never wired into the Python API.
if backend == "ollama":
    num_ctx = int(os.environ.get("GRAPHIFY_OLLAMA_NUM_CTX", 32768))
    kwargs["extra_body"] = {"options": {"num_ctx": num_ctx}}

resp = client.chat.completions.create(**kwargs)