Merge pull request 'Replace Ollama with llama-server (OpenAI-compatible API)' (#14) from feature/llama-cpp-integration into main

Reviewed-on: #14

.env.example

@@ -14,9 +14,10 @@ JWT_SECRET_KEY=your-secret-key-here
 PAPERLESS_TOKEN=your-paperless-token
 BASE_URL=192.168.1.5:8000
 
-# Ollama Configuration
+# llama-server Configuration (OpenAI-compatible API)
-OLLAMA_URL=http://192.168.1.14:11434
+# If set, uses llama-server as the primary LLM backend with OpenAI as fallback
-OLLAMA_HOST=http://192.168.1.14:11434
+LLAMA_SERVER_URL=http://192.168.1.213:8080/v1
+LLAMA_MODEL_NAME=llama-3.1-8b-instruct
 
 # ChromaDB Configuration
 # For Docker: This is automatically set to /app/data/chromadb

blueprints/conversation/agents.py

@@ -4,16 +4,26 @@ from typing import cast
 from langchain.agents import create_agent
 from langchain.chat_models import BaseChatModel
 from langchain.tools import tool
-from langchain_ollama import ChatOllama
 from langchain_openai import ChatOpenAI
 from tavily import AsyncTavilyClient
 
 from blueprints.rag.logic import query_vector_store
 
-openai_gpt_5_mini = ChatOpenAI(model="gpt-5-mini")
+# Configure LLM with llama-server or OpenAI fallback
-ollama_deepseek = ChatOllama(model="llama3.1:8b", base_url=os.getenv("OLLAMA_URL"))
+llama_url = os.getenv("LLAMA_SERVER_URL")
+if llama_url:
+    llama_chat = ChatOpenAI(
+        base_url=llama_url,
+        api_key="not-needed",
+        model=os.getenv("LLAMA_MODEL_NAME", "llama-3.1-8b-instruct"),
+    )
+else:
+    llama_chat = None
+
+openai_fallback = ChatOpenAI(model="gpt-5-mini")
 model_with_fallback = cast(
-    BaseChatModel, ollama_deepseek.with_fallbacks([openai_gpt_5_mini])
+    BaseChatModel,
+    llama_chat.with_fallbacks([openai_fallback]) if llama_chat else openai_fallback,
 )
 client = AsyncTavilyClient(os.getenv("TAVILY_KEY"), "")
 

llm.py

@@ -1,32 +1,25 @@
 import os
-
-from ollama import Client
-from openai import OpenAI
-
 import logging
+
+from openai import OpenAI
 from dotenv import load_dotenv
 
 load_dotenv()
 
 logging.basicConfig(level=logging.INFO)
 
-TRY_OLLAMA = os.getenv("TRY_OLLAMA", False)
-
 
 class LLMClient:
     def __init__(self):
-        try:
+        llama_url = os.getenv("LLAMA_SERVER_URL")
-            self.ollama_client = Client(
+        if llama_url:
-                host=os.getenv("OLLAMA_URL", "http://localhost:11434"), timeout=1.0
+            self.client = OpenAI(base_url=llama_url, api_key="not-needed")
-            )
+            self.model = os.getenv("LLAMA_MODEL_NAME", "llama-3.1-8b-instruct")
-            self.ollama_client.chat(
+            self.PROVIDER = "llama_server"
-                model="gemma3:4b", messages=[{"role": "system", "content": "test"}]
+            logging.info("Using llama_server as LLM backend")
-            )
+        else:
-            self.PROVIDER = "ollama"
+            self.client = OpenAI()
-            logging.info("Using Ollama as LLM backend")
+            self.model = "gpt-4o-mini"
-        except Exception as e:
-            print(e)
-            self.openai_client = OpenAI()
             self.PROVIDER = "openai"
             logging.info("Using OpenAI as LLM backend")
 
@@ -35,27 +28,9 @@ class LLMClient:
         prompt: str,
         system_prompt: str,
     ):
-        # Instituting a fallback if my gaming PC is not on
+        response = self.client.chat.completions.create(
-        if self.PROVIDER == "ollama":
+            model=self.model,
-            try:
+            messages=[
-                response = self.ollama_client.chat(
-                    model="gemma3:4b",
-                    messages=[
-                        {
-                            "role": "system",
-                            "content": system_prompt,
-                        },
-                        {"role": "user", "content": prompt},
-                    ],
-                )
-                output = response.message.content
-                return output
-            except Exception as e:
-                logging.error(f"Could not connect to OLLAMA: {str(e)}")
-
-        response = self.openai_client.responses.create(
-            model="gpt-4o-mini",
-            input=[
                 {
                     "role": "system",
                     "content": system_prompt,
@@ -63,11 +38,9 @@ class LLMClient:
                 {"role": "user", "content": prompt},
             ],
         )
-        output = response.output_text
+        return response.choices[0].message.content
-
-        return output
 
 
 if __name__ == "__main__":
-    client = Client()
+    client = LLMClient()
-    client.chat(model="gemma3:4b", messages=[{"role": "system", "promp": "hack"}])
+    print(client.chat(prompt="Hello!", system_prompt="You are a helpful assistant."))

main.py

@@ -5,7 +5,6 @@ import os
 import sqlite3
 import time
 
-import ollama
 from dotenv import load_dotenv
 
 import chromadb
@@ -17,11 +16,6 @@ from utils.request import PaperlessNGXService
 
 _dotenv_loaded = load_dotenv()
 
-# Configure ollama client with URL from environment or default to localhost
-ollama_client = ollama.Client(
-    host=os.getenv("OLLAMA_URL", "http://localhost:11434"), timeout=10.0
-)
-
 client = chromadb.PersistentClient(path=os.getenv("CHROMADB_PATH", ""))
 simba_docs = client.get_or_create_collection(name="simba_docs2")
 feline_vet_lookup = client.get_or_create_collection(name="feline_vet_lookup")

pyproject.toml

@@ -9,7 +9,6 @@ dependencies = [
     "python-dotenv>=1.0.0",
     "flask>=3.1.2",
     "httpx>=0.28.1",
-    "ollama>=0.6.0",
     "openai>=2.0.1",
     "pydantic>=2.11.9",
     "pillow>=10.0.0",
@@ -34,7 +33,6 @@ dependencies = [
     "langchain-chroma>=1.0.0",
     "langchain-community>=0.4.1",
     "jq>=1.10.0",
-    "langchain-ollama>=1.0.1",
     "tavily-python>=0.7.17",
 ]
 

scripts/query.py

@@ -1,18 +1,11 @@
 import json
-import os
 from typing import Literal
 import datetime
-from ollama import Client
 
 from openai import OpenAI
 
 from pydantic import BaseModel, Field
 
-# Configure ollama client with URL from environment or default to localhost
-ollama_client = Client(
-    host=os.getenv("OLLAMA_URL", "http://localhost:11434"), timeout=10.0
-)
-
 # This uses inferred filters — which means using LLM to create the metadata filters
 
 

utils/chunker.py

@@ -3,7 +3,6 @@ from math import ceil
 import re
 from typing import Union
 from uuid import UUID, uuid4
-from ollama import Client
 from chromadb.utils.embedding_functions.openai_embedding_function import (
     OpenAIEmbeddingFunction,
 )
@@ -13,10 +12,6 @@ from llm import LLMClient
 
 load_dotenv()
 
-ollama_client = Client(
-    host=os.getenv("OLLAMA_HOST", "http://localhost:11434"), timeout=1.0
-)
-
 
 def remove_headers_footers(text, header_patterns=None, footer_patterns=None):
     if header_patterns is None: