I want to create an LLM that can call custom-made APIs (note there are many APIs that we have already created). The LLM can make all the HTTP requests(get, post, put, delete). It can infer which llm to call and also call it with the required parameters. So that the user can just use natural language to do tasks. For instance, we have a budget api

 def budget(db, request):
    
    # Retrieve data from the request JSON
    data_db = request.json

    try:
        collection = db["budget"]

        # Handle inserting budget
        if request.method == 'POST':
            # Add creation_date to the data
            data_db['creation_date'] = datetime.now().strftime('%Y-%m-%d %H:%M')
            # Add balance to the data
            data_db['balance'] = 0.0
            # Add budget_project_type name
            data_db['budget_project_type_name'] = db["budget_project_type"].find_one({"_id": ObjectId(data_db['budget_project_type'])})['name']
            # Add budget_group name
            data_db['budget_group_name'] = db["budget_group"].find_one({"_id": ObjectId(data_db['budget_group'])})['name']
            # Add blank dicts
            data_db['expenses'], data_db['revenue'] = {}, {}
            # Insert the budget data into the database
            collection.insert_one(data_db)
            return {"budget_insert" : True}
        
        # Handle getting budget
        elif request.method == 'GET':
            if data_db['budget_get'] == 'unique':
                response = collection.find_one({"_id": ObjectId(data_db['_id']), "email": data_db['email']})
                response['_id'] = str(response['_id'])
            elif data_db['budget_get'] == 'all':
                response = collection.find({"email": data_db['email']})
                response_list = []
                for result in list(response):
                    result['_id'] = str(result['_id'])
                    response_list.append(result)
                response = response_list
            return jsonify(response)
            
        # Handle budget update
        elif request.method == 'PUT':
            collection.update_one({"$and": [{"_id": ObjectId(data_db['_id'])}, {"email": data_db['email']}]}, {"$set": data_db})
            return {"budget_update" : True}

        # Handle budget deletion
        elif request.method == 'DELETE':
            collection.delete_one({"$and": [{"_id": ObjectId(data_db['_id']), "email": data_db['email']}]})
            return {"budget_delete" : True}

        # Handle unknown methods
        else:
            return {"error" : f"budget_{request.method}_unsupported"}
    except Exception as e:
        print(e)
        return {"error" : "except_budget_module"}
`

@app.route('/budget', methods=['POST', 'GET', 'PUT', 'DELETE'])
def budget_():
    return budget(db, request)

I looked into Openai’s function calling I believe that it’s really expensive with more number of functions.
Their documentation suggests fine-tuning the model to save tokens to which Im confused on how this would help in my case.
I have experience using LLamaindex but I want to use langchain for this project as I’ve seen better documentation regarding API calling.

How do I proceed with this project? I would prefer using a local method as Im using Ollama.Is function calling the best option I should look into? Or should I just use OpenAI’s function calling since local models like llama2 7b aren’t that good?