ollama-server: Gap analysis — missing endpoints & fields for drop-in compatibility

[adiled]

ollama-server: Gap Analysis — Missing Endpoints & Fields

Goal: ollama-server should be a drop-in replacement for Ollama's API surface so
humd can sit on port 11434 and any Ollama client (open-webui, cline, langchain, etc.)
works without changes.

Current endpoints

Endpoint	Status
GET /	✅
GET /api/tags	✅ (synthesized)
POST /api/chat	✅
POST /api/generate	✅

Missing core endpoints

#	Endpoint	Method	Description	Priority
1	/api/version	GET	Returns {"version": "0.5.1"} — some clients check this	🔴 High
2	/api/show	POST	Model info: modelfile, parameters, template, details, model_info, capabilities — critical for RAG/agents that inspect models	🔴 High
3	/api/pull	POST	Pull/download a model from Ollama library — streaming progress, resumable	🟡 Medium
4	/api/push	POST	Push a model to a model library — streaming progress	🟡 Medium
5	/api/create	POST	Create a model from another model, safetensors dir, or GGUF file — streaming progress	🟡 Medium
6	HEAD /api/blobs/:digest	HEAD	Check if a blob exists on the server	🟡 Medium
7	POST /api/blobs/:digest	POST	Push a file blob to the server (used by /api/create)	🟡 Medium
8	/api/copy	POST	Copy/rename a model (source → destination)	🟢 Low
9	/api/delete	DELETE	Delete a model and its data	🟢 Low
10	/api/ps	GET	List currently loaded-in-memory models (with expires_at, size_vram)	🔴 High
11	/api/embed	POST	Generate embeddings from a model — the main RAG/embedding endpoint	🔴 High
12	/api/embeddings	POST	Deprecated endpoint but still used by older clients	🟢 Low

Missing request/response fields on existing endpoints

POST /api/chat — Missing request fields

Field	Description	Priority
think	For thinking models: should the model think before responding?	🔴 High
format	json or JSON schema for structured outputs	🔴 High
keep_alive	Keeps model loaded (0 = unload, 5m = default, never = stay)	🔴 High
images	Base64-encoded images in messages for multimodal	🟡 Medium
messages[].tool_name	Tool call result with the tool name	🟡 Medium
messages[].tool_calls	Client sends back tool call results in history	🟡 Medium
options (full schema)	num_keep, seed, num_predict, top_k, top_p, temperature, stop, etc.	🟡 Medium

POST /api/chat — Missing response fields

Field	Description	Priority
message.thinking	Model's thinking process for thinking models	🔴 High
message.tool_name	Present when model returns a tool call	🟡 Medium
total_duration	Nanoseconds generating response	🟡 Medium
load_duration	Nanoseconds loading model	🟡 Medium
prompt_eval_count	Token count in prompt	🟡 Medium
prompt_eval_duration	Nanoseconds evaluating prompt	🟡 Medium
eval_count	Token count in response	🟡 Medium
eval_duration	Nanoseconds generating response	🟡 Medium

POST /api/generate — Additional missing fields

Field	Description	Priority
suffix	Text after the model response	🟡 Medium
raw	Bypass prompt templating	🟡 Medium

Missing: Model management concept

Current ollama-server only has a static list of models (OLLAMA_SERVER_MODELS). Needs:

1. Dynamic model registry
2. /api/show with real metadata per model
3. Model lifecycle awareness via keep_alive

Priority summary

P0 (must-have for drop-in compatibility):

• GET /api/version, POST /api/show, GET /api/ps
• think / response.thinking
• format (JSON/structured output)
• keep_alive handling
• Response time metrics

P1 (important for broader client support):

• POST /api/embed for embeddings
• suffix, raw, images, options.* on generate/chat

P2 (nice to have):

• Model management: pull, push, create, blobs, copy, delete