Fix/configs and llm resp selection (#476)

* fix: rename goals prompt

* fix: send hyps to llm based resp selector

* fix: send configs and default values

* fix: configs

* fix: extra file

* fix: extra imports

common/prompts.py

@@ -8,7 +8,7 @@
 logger = logging.getLogger(__name__)
 
 with open("common/prompts/goals_for_prompts.json", "r") as f:
-    META_PROMPT = json.load(f)["prompt"]
+    META_GOALS_PROMPT = json.load(f)["prompt"]
 
 
 def send_request_to_prompted_generative_service(dialog_context, prompt, url, config, timeout, sending_variables):

response_selectors/llm_based_response_selector/server.py

@@ -3,13 +3,13 @@
 import json
 import logging
 import numpy as np
-import requests
 import time
 from copy import deepcopy
 from os import getenv
 
 import sentry_sdk
 from flask import Flask, request, jsonify
+from common.prompts import send_request_to_prompted_generative_service
 from common.utils import is_toxic_or_badlisted_utterance
 
 
@@ -62,18 +62,15 @@ def select_response_by_scores(hypotheses, scores):
 
 def select_response(dialog_context, hypotheses):
     try:
-        response = requests.post(
+        response = send_request_to_prompted_generative_service(
+            dialog_context,
+            PROMPT + "\nHypotheses:\n" + "\n".join([f'"{hyp["text"]}"' for hyp in hypotheses]),
             GENERATIVE_SERVICE_URL,
-            json={
-                "dialog_contexts": [dialog_context],
-                "prompts": [PROMPT],
-                "configs": [GENERATIVE_SERVICE_CONFIG],
-                **sending_variables,
-            },
-            timeout=GENERATIVE_TIMEOUT,
+            GENERATIVE_SERVICE_CONFIG,
+            GENERATIVE_TIMEOUT,
+            sending_variables,
         )
-        # batch of a list of one string [["this is the response"]]
-        result = response.json()[0][0]
+        result = response[0]
     except Exception as e:
         sentry_sdk.capture_exception(e)
         logger.exception(e)
@@ -114,6 +111,7 @@ def respond():
             selected_bot_attributes.append(hypotheses[best_id].pop("bot_attributes", {}))
             hypotheses[best_id].pop("annotations", {})
             selected_attributes.append(hypotheses[best_id])
+
         except Exception as e:
             sentry_sdk.capture_exception(e)
             logger.exception(e)
@@ -122,6 +120,7 @@ def respond():
                 "Selected a response with the highest confidence."
             )
             selected_resp, best_id = select_response_by_scores(hypotheses, [hyp["confidence"] for hyp in hypotheses])
+
             selected_responses.append(hypotheses[best_id].pop("text"))
             selected_skill_names.append(hypotheses[best_id].pop("skill_name"))
             selected_confidences.append(hypotheses[best_id].pop("confidence"))

services/openai_api_lm/server.py

@@ -5,7 +5,7 @@
 
 import openai
 import sentry_sdk
-from common.prompts import META_PROMPT
+from common.prompts import META_GOALS_PROMPT
 from common.universal_templates import GENERATIVE_ROBOT_TEMPLATE
 from flask import Flask, request, jsonify
 from sentry_sdk.integrations.flask import FlaskIntegration
@@ -96,7 +96,8 @@ def respond():
     st_time = time.time()
     contexts = request.json.get("dialog_contexts", [])
     prompts = request.json.get("prompts", [])
-    configs = request.json.get("configs", [])
+    configs = request.json.get("configs", None)
+    configs = [None] * len(prompts) if configs is None else configs
     configs = [DEFAULT_CONFIGS[PRETRAINED_MODEL_NAME_OR_PATH] if el is None else el for el in configs]
     if len(contexts) > 0 and len(prompts) == 0:
         prompts = [""] * len(contexts)
@@ -144,7 +145,7 @@ def generate_goals():
     try:
         responses = []
         for openai_api_key, openai_org, prompt, config in zip(openai_api_keys, openai_orgs, prompts, configs):
-            context = ["hi", META_PROMPT + f"\nPrompt: '''{prompt}'''\nResult:"]
+            context = ["hi", META_GOALS_PROMPT + f"\nPrompt: '''{prompt}'''\nResult:"]
             goals_for_prompt = generate_responses(context, openai_api_key, openai_org, "", config)[0]
             logger.info(f"Generated goals: `{goals_for_prompt}` for prompt: `{prompt}`")
             responses += [goals_for_prompt]

services/transformers_lm/server.py

@@ -9,7 +9,7 @@
 from sentry_sdk.integrations.flask import FlaskIntegration
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-from common.prompts import META_PROMPT
+from common.prompts import META_GOALS_PROMPT
 from common.universal_templates import GENERATIVE_ROBOT_TEMPLATE
 
 
@@ -120,7 +120,8 @@ def respond():
     st_time = time.time()
     contexts = request.json.get("dialog_contexts", [])
     prompts = request.json.get("prompts", [])
-    configs = request.json.get("configs", [])
+    configs = request.json.get("configs", None)
+    configs = [None] * len(prompts) if configs is None else configs
     configs = [DEFAULT_CONFIGS[PRETRAINED_MODEL_NAME_OR_PATH] if el is None else el for el in configs]
     if len(contexts) > 0 and len(prompts) == 0:
         prompts = [""] * len(contexts)
@@ -161,7 +162,7 @@ def generate_goals():
     try:
         responses = []
         for prompt, config in zip(prompts, configs):
-            context = ["hi", META_PROMPT + f"\nPrompt: '''{prompt}'''\nResult:"]
+            context = ["hi", META_GOALS_PROMPT + f"\nPrompt: '''{prompt}'''\nResult:"]
             goals_for_prompt = generate_responses(context, model, tokenizer, "", config)[0]
             logger.info(f"Generated goals: `{goals_for_prompt}` for prompt: `{prompt}`")
             responses += [goals_for_prompt]

services/transformers_peft_lm/server.py

@@ -10,7 +10,7 @@
 from sentry_sdk.integrations.flask import FlaskIntegration
 from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
 
-from common.prompts import META_PROMPT
+from common.prompts import META_GOALS_PROMPT
 from common.universal_templates import GENERATIVE_ROBOT_TEMPLATE
 
 
@@ -146,7 +146,7 @@ def generate_goals():
     try:
         responses = []
         for prompt in prompts:
-            context = ["hi", META_PROMPT + f"\nPrompt: '''{prompt}'''\nResult:"]
+            context = ["hi", META_GOALS_PROMPT + f"\nPrompt: '''{prompt}'''\nResult:"]
             goals_for_prompt = generate_responses(context, model, tokenizer, "")[0]
             logger.info(f"Generated goals: `{goals_for_prompt}` for prompt: `{prompt}`")
             responses += [goals_for_prompt]