Feat/universal skill selector (#13)

* feat: universal resp selector for debug * feat: universal resp selector for debug * feat: universal resp selector distribution * fix; upd services lm * feat: universal skill_selector * feat: universal skill working with multiple prompts and llms * feat: component card * fix; getting skill name not none * fix: codestyle * fix: resp selector vars * fix: remove sending apikeys with raw uttr attrs * fix: increase timeout * fix: do not use user attrs * fix: promtp to skill seletor * fix: select all skills (for resp selector debug) * fix: distribution with skill selector * fix: attrs * fix: attrs extra * fix: turning on all skills and getting skill names from input * fix: default value * feat: return all hypotheses for debug mode * fix: order of calls * fix: remove external internal * feat: docs * feat: component cards * fix: env in cards --------- Co-authored-by: smilni <[email protected]>
deeppavlov · Sep 6, 2023 · 5f8133a · 5f8133a
1 parent f4779b8
commit 5f8133a
Show file tree

Hide file tree

Showing 18 changed files with 435 additions and 11 deletions.
diff --git a/assistant_dists/universal_selectors_assistant/dev.yml b/assistant_dists/universal_selectors_assistant/dev.yml
@@ -10,6 +10,13 @@ services:
       - "./annotators/SentSeg:/src"
     ports:
       - 8011:8011
+  universal-llm-based-skill-selector:
+    volumes:
+      - "./skill_selectors/universal_llm_based_skill_selector:/src"
+      - "./common:/src/common"
+      - "./components:/src/components"
+    ports:
+      - 8187:8187
   universal-llm-based-response-selector:
     volumes:
       - "./response_selectors/universal_llm_based_response_selector:/src"

diff --git a/assistant_dists/universal_selectors_assistant/docker-compose.override.yml b/assistant_dists/universal_selectors_assistant/docker-compose.override.yml
@@ -3,7 +3,7 @@ services:
     command: sh -c 'bin/wait && python -m deeppavlov_agent.run agent.pipeline_config=assistant_dists/universal_selectors_assistant/pipeline_conf.json'
     environment:
       WAIT_HOSTS: "sentseg:8011, universal-llm-based-response-selector:8181, combined-classification:8087, 
-        sentence-ranker:8128, openai-api-chatgpt:8145, openai-api-davinci3:8131,
+        sentence-ranker:8128, openai-api-chatgpt:8145, openai-api-davinci3:8131, universal-llm-based-skill-selector:8187,
         openai-api-gpt4:8159, openai-api-gpt4-32k:8160, openai-api-chatgpt-16k:8167,
         dff-universal-prompted-skill:8147"
       WAIT_HOSTS_TIMEOUT: ${WAIT_TIMEOUT:-1000}
@@ -28,8 +28,9 @@ services:
     env_file: [ .env ]
     build:
       args:
-        CONFIG: combined_classifier.json
         SERVICE_PORT: 8087
+        SERVICE_NAME: combined_classification
+        CONFIG: combined_classifier.json
       context: .
       dockerfile: ./annotators/combined_classification/Dockerfile
     command: gunicorn --workers=1 server:app -b 0.0.0.0:8087 --timeout 600
@@ -42,6 +43,29 @@ services:
         reservations:
           memory: 2G
 
+  universal-llm-based-skill-selector:
+    env_file: [ .env ]
+    build:
+      args:
+        SERVICE_PORT: 8187
+        SERVICE_NAME: skill_selector
+        DEFAULT_LM_SERVICE_URL: http://openai-api-chatgpt:8145/respond
+        DEFAULT_LM_SERVICE_CONFIG: openai-chatgpt.json
+        GENERATIVE_TIMEOUT: 20
+        N_UTTERANCES_CONTEXT: 3
+        MAX_N_SKILLS: 3
+      context: .
+      dockerfile: ./skill_selectors/universal_llm_based_skill_selector/Dockerfile
+    command: flask run -h 0.0.0.0 -p 8187
+    environment:
+      - FLASK_APP=server
+    deploy:
+      resources:
+        limits:
+          memory: 100M
+        reservations:
+          memory: 100M
+
   universal-llm-based-response-selector:
     env_file: [ .env ]
     build:

diff --git a/assistant_dists/universal_selectors_assistant/pipeline_conf.json b/assistant_dists/universal_selectors_assistant/pipeline_conf.json
@@ -112,12 +112,13 @@
             }
         },
         "skill_selectors": {
-            "rule_based_selector": {
+            "skill_selector": {
                 "connector": {
-                    "protocol": "python",
-                    "class_name": "skill_selectors.rule_based_selector.connector:RuleBasedSkillSelectorConnector"
+                    "protocol": "http",
+                    "timeout": 10.0,
+                    "url": "http://universal-llm-based-skill-selector:8187/respond"
                 },
-                "dialog_formatter": "state_formatters.dp_formatters:base_skill_selector_formatter_dialog",
+                "dialog_formatter": "state_formatters.dp_formatters:cropped_dialog",
                 "response_formatter": "state_formatters.dp_formatters:simple_formatter_service",
                 "previous_services": [
                     "annotators"
@@ -127,8 +128,8 @@
                 ],
                 "is_enabled": true,
                 "source": {
-                    "component": "components/xSwFvtAUdvtQosvzpb7oMg.yml",
-                    "service": "skill_selectors/rule_based_selector/service_configs/agent"
+                    "component": "components/jkdfh12oien9dfuih1.yml",
+                    "service": "skill_selectors/universal_llm_based_skill_selector/service_configs/universal-llm-based-skill-selector"
                 }
             }
         },

diff --git a/components.tsv b/components.tsv
@@ -186,4 +186,10 @@
 8182    llm-based-skill-selector
 8183	external-integration-skill
 8184	external-fake-server
-8185    dff-roles-prompted-skill
+8185	dff-roles-prompted-skill
+8186
+8187	universal-llm-based-skill-selector
+8188
+8189
+8190
+8191
diff --git a/components/jkdfh12oien9dfuih1.yml b/components/jkdfh12oien9dfuih1.yml
@@ -0,0 +1,25 @@
+name: skill_selector
+display_name: Universal LLM-based Skill Selector
+component_type: null
+model_type: Dictionary/Pattern-based
+is_customizable: false
+author: [email protected]
+description: Algorithm that selects skills to generate hypotheses among the given list of all available
+  skills in pipeline and provided LLM service
+ram_usage: 100M
+gpu_usage: null
+group: skill_selectors
+connector:
+  protocol: http
+  timeout: 10.0
+  url: http://universal-llm-based-skill-selector:8187/respond
+dialog_formatter: state_formatters.dp_formatters:cropped_dialog
+response_formatter: state_formatters.dp_formatters:simple_formatter_service
+previous_services:
+- annotators
+required_previous_services: null
+state_manager_method: null
+tags: null
+endpoint: respond
+service: skill_selectors/universal_llm_based_skill_selector/service_configs/universal-llm-based-skill-selector
+date_created: '2023-03-16T09:45:32'
diff --git a/response_selectors/llm_based_response_selector/server.py b/response_selectors/llm_based_response_selector/server.py
@@ -58,6 +58,9 @@ def select_response_by_scores(hypotheses, scores):
 
 
 def select_response(dialog_context, hypotheses, human_uttr_attributes):
+    if human_uttr_attributes.get("return_all_hypotheses", None):
+        return "\n".join(['"' + hyp["skill_name"] + '": "' + hyp["text"] + '"' for hyp in hypotheses])
+
     try:
         ie_types = [
             "external service" if hyp["skill_name"] in EXTERNAL_SKILLS else "internal service" for hyp in hypotheses
@@ -85,6 +88,7 @@ def select_response(dialog_context, hypotheses, human_uttr_attributes):
             sending_variables,
         )
         result = response[0]
+        result = result.replace("[internal service]", "").replace("[external service]", "").strip()
     except Exception as e:
         sentry_sdk.capture_exception(e)
         logger.exception(e)

diff --git a/response_selectors/universal_llm_based_response_selector/README.md b/response_selectors/universal_llm_based_response_selector/README.md
@@ -45,7 +45,7 @@ for lm_service in ["ChatGPT"]:
                 {"openai_api_key": "FILL IN"}
             ],
             # ---------------------------- response selector parameters
-            "response_selector_prompt": "Select the most funny answer.", 
+            "response_selector_prompt": "Select the most funny answer.\nLIST_OF_HYPOTHESES\n", 
             "response_selector_lm_service_url": LM_SERVICES_MAPPING[lm_service],
             "response_selector_lm_service_config": 
                 {

diff --git a/response_selectors/universal_llm_based_response_selector/server.py b/response_selectors/universal_llm_based_response_selector/server.py
@@ -71,6 +71,9 @@ def select_response_by_scores(hypotheses, scores):
 def select_response(dialog, hypotheses, human_uttr_attributes):
     dialog_context = [uttr["text"] for uttr in dialog["utterances"][-N_UTTERANCES_CONTEXT:]]
 
+    if human_uttr_attributes.get("return_all_hypotheses", None):
+        return "\n".join(['"' + hyp["skill_name"] + '": "' + hyp["text"] + '"' for hyp in hypotheses])
+
     # get prompt from the current utterance attributes
     given_prompt = human_uttr_attributes.get("response_selector_prompt", DEFAULT_PROMPT)
     for i in range(1, len(dialog["utterances"]) + 1, 2):
@@ -113,6 +116,7 @@ def select_response(dialog, hypotheses, human_uttr_attributes):
             sending_variables,
         )
         result = response[0]
+        result = result.replace("[internal service]", "").replace("[external service]", "").strip()
     except Exception as e:
         sentry_sdk.capture_exception(e)
         logger.exception(e)

diff --git a/...based_response_selector/service_configs/universal-llm-based-response-selector/service.yml b/...based_response_selector/service_configs/universal-llm-based-response-selector/service.yml
@@ -2,7 +2,8 @@ name: universal-llm-based-response-selector
 endpoints:
 - respond
 compose:
-  env_file: [ .env,.env_secret ]
+  env_file:
+  - .env
   build:
     args:
       SERVICE_PORT: 8181

diff --git a/skill_selectors/llm_based_skill_selector/server.py b/skill_selectors/llm_based_skill_selector/server.py
@@ -58,6 +58,9 @@ def select_skills(dialog):
     # pipeline is smth like this: ['annotators.sentseg', 'skills.dummy_skill',
     # 'candidate_annotators.sentence_ranker', 'response_selectors.response_selector', ...]
     all_skill_names = [el.split(".")[1] for el in pipeline if "skills" in el]
+    if human_uttr_attributes.get("selected_skills", None) in ["all", []]:
+        logger.info(f"llm_based_skill_selector selected ALL skills:\n`{all_skill_names}`")
+        return all_skill_names
 
     try:
         if "LIST_OF_AVAILABLE_AGENTS_WITH_DESCRIPTIONS" in PROMPT:

diff --git a/skill_selectors/universal_llm_based_skill_selector/Dockerfile b/skill_selectors/universal_llm_based_skill_selector/Dockerfile
@@ -0,0 +1,29 @@
+FROM python:3.10
+
+RUN mkdir /src
+
+COPY ./skill_selectors/universal_llm_based_skill_selector/requirements.txt /src/requirements.txt
+RUN pip install -r /src/requirements.txt
+
+ARG SERVICE_PORT
+ENV SERVICE_PORT ${SERVICE_PORT}
+ARG GENERATIVE_TIMEOUT
+ENV GENERATIVE_TIMEOUT ${GENERATIVE_TIMEOUT}
+ARG ENVVARS_TO_SEND
+ENV ENVVARS_TO_SEND ${ENVVARS_TO_SEND}
+ARG N_UTTERANCES_CONTEXT=5
+ENV N_UTTERANCES_CONTEXT ${N_UTTERANCES_CONTEXT}
+ARG DEFAULT_LM_SERVICE_URL
+ENV DEFAULT_LM_SERVICE_URL ${DEFAULT_LM_SERVICE_URL}
+ARG DEFAULT_LM_SERVICE_CONFIG
+ENV DEFAULT_LM_SERVICE_CONFIG ${DEFAULT_LM_SERVICE_CONFIG}
+ARG MAX_N_SKILLS
+ENV MAX_N_SKILLS ${MAX_N_SKILLS}
+
+COPY skill_selectors/universal_llm_based_skill_selector /src
+WORKDIR /src
+COPY ./common/ ./common/
+COPY ./components/ ./components/
+
+
+CMD gunicorn --workers=1 server:app -b 0.0.0.0:${SERVICE_PORT} --timeout=1200
diff --git a/skill_selectors/universal_llm_based_skill_selector/README.md b/skill_selectors/universal_llm_based_skill_selector/README.md
@@ -0,0 +1,92 @@
+# Universal LLM-based Skill Selector
+
+## Description
+
+Debugging Skill Selector is a component selecting a list of skills to generate hypotheses.
+The LLM-based Skill Selector utilizes LLM service to select the skills in a generative manner.
+The considered LLM-service, generative parameters and prompt  are provided IN attributes of 
+the last human utterance. The list of all available skills is picked up from human utterance attributes.
+
+
+**Important** Provide `"return_all_hypotheses": True` (to return joined list of all returned hypotheses) 
+and `"selected_skills": "all"` (to turn on dff_universal_prompted_skill because all other prompted skills
+are not deployed during debugging process).
+
+How to use:
+
+```python
+import requests
+import random
+from time import sleep
+
+
+LM_SERVICES_MAPPING = {
+    "ChatGPT": "http://openai-api-chatgpt:8145/respond",
+}
+for lm_service in ["ChatGPT"]:
+    print(f"Checking `Universal Assistant` with `{lm_service}`")
+
+    result = requests.post(
+        "http://0.0.0.0:4242", 
+        json={
+            "user_id": f"test-user-{random.randint(100, 1000)}", 
+            "payload": "How much is two plus two?",
+            # ---------------------------- batch of universal skills to generate hypotheses
+            "skill_name": ["Mathematician Skill", "blondy_skill"],
+            "prompt": ["Answer as a mathematician.", "Answer like you are a stupid Blondy Girl."], 
+            "lm_service_url": [LM_SERVICES_MAPPING[lm_service], LM_SERVICES_MAPPING[lm_service]],
+            "lm_service_config": [
+                {
+                    "max_new_tokens": 64,
+                    "temperature": 0.9,
+                    "top_p": 1.0,
+                    "frequency_penalty": 0,
+                    "presence_penalty": 0
+                }, 
+                None
+            ],
+            "lm_service_kwargs": [
+                {"openai_api_key": "FILL IN"},
+                {"openai_api_key": "FILL IN"}
+            ],
+            # ---------------------------- response selector MUST RETURN ALL HYPOTHESES JOINED
+            "return_all_hypotheses": True,
+            # ---------------------------- skill selector 
+            "selected_skills": "all", # must use it to turn on universal skill (others are not deployed!)
+            "skill_selector_prompt": """
+Select up to 2 the most relevant to the dialog context skills based on the given short descriptions of abilities of different skills of the assistant.
+
+Skills:
+"Mathematician Skill": "A skill pretending to be a mathematician."
+"blondy_skill": "A Skill pretending to be a blondy girl."
+
+Return only names of the selected skills divided by comma. Do not respond to the dialog context.""", 
+            "skill_selector_lm_service_url": LM_SERVICES_MAPPING[lm_service],
+            "skill_selector_lm_service_config": 
+                {
+                    "max_new_tokens": 64,
+                    "temperature": 0.4,
+                    "top_p": 1.0,
+                    "frequency_penalty": 0,
+                    "presence_penalty": 0
+                },
+            "skill_selector_lm_service_kwargs": {"openai_api_key": "FILL IN"},
+        }).json()
+    print(f"Response:\n{result['response']}")
+    if result["active_skill"] not in ["Dummy Skill", "dummy_skill"]:
+        print("Success!")
+    elif lm_service in ["ChatGPT", "GPT-3.5"]:
+        print(f"\nATTENTION! OpenAI services do not work!\n")
+    else:
+        print(f"\nERROR: `Universal Assistant` returned `{result}`\n")
+    sleep(5)
+```
+
+### Parameters
+
+The algorithm utilizes `N_UTTERANCES_CONTEXT` last utterances as a context for LLM.
+Number of returned skills can ve varied by the prompt itself.
+
+## Dependencies
+
+- generative service `DEFAULT_LM_SERVICE_URL` or the given in attributes in the last human utterance
diff --git a/skill_selectors/universal_llm_based_skill_selector/requirements.txt b/skill_selectors/universal_llm_based_skill_selector/requirements.txt
@@ -0,0 +1,9 @@
+flask==1.1.1
+itsdangerous==2.0.1
+gunicorn==19.9.0
+requests==2.22.0
+numpy==1.25.0
+sentry-sdk==0.12.3
+jinja2<=3.0.3
+Werkzeug<=2.0.3
+pyyaml==6.0.1