Merge pull request #520 from deeppavlov/dev

Release v1.8.2

annotators/doc_retriever/README.md

@@ -0,0 +1,22 @@
+# Document Retriever
+
+## Description
+
+Document Retriever is an annotator with two endpoints used to retrieve `PARAGRAPHS_NUM` document parts most relevant to the user request.
+
+1. **train_and_upload_model** endpoint converts the documents provided by the user to txt format (if necessary) and splits them into chunks of ~100 words. Chunks are then transformed into a TF-IDF matrix; the resulting vectors and the vectorizer are saved for future use. This step is performed only once, in the beginning of the dialog.
+Documents (txt format), matrix, and vectorizer are uploaded to file server to be used by **return_candidates** endpoint and **dff_document_qa_llm** skill.
+2. **return_candidates** endpoint downloads TF-IDF matrix and vectorizer from the file server. It then converts the user’s utterance into a TF-IDF vector and finds `PARAGRAPHS_NUM` candidates with highest cosine similarity among TF-IDF vectors of text chunks.
+
+## Parameters
+
+```
+CONFIG_PATH: configuration file with parameters for doc_retriever model
+FILE_SERVER_TIMEOUT: timeout for request where files are stored
+PARAGRAPHS_NUM: number of most relevant chunks to retrieve. Don't make this number too large or the chunks won't fit into LLM context!
+DOC_PATH_OR_LINK: paths or link to the files to be use for Question Answering. If paths, those are paths to files in `documents` folder in dream. If links, those must point to a file, not an Internet page. NB: file paths/links must be separated by a comma and no whitespace. 
+```
+
+## Dependencies
+
+- **return_candidates** endpoint depends on **train_and_upload_model** endpoint

annotators/prompt_selector/service_configs/dream_persona_openai_prompted/environment.yml

@@ -1,5 +1,5 @@
 SERVICE_PORT: 8135
 SERVICE_NAME: prompt_selector
 N_SENTENCES_TO_RETURN: 3
-PROMPTS_TO_CONSIDER: dream_persona
+PROMPTS_TO_CONSIDER: dream_persona,dream_faq
 FLASK_APP: server

annotators/prompt_selector/service_configs/dream_persona_openai_prompted/service.yml

@@ -10,7 +10,7 @@ compose:
       SERVICE_PORT: 8135
       SERVICE_NAME: prompt_selector
       N_SENTENCES_TO_RETURN: 3
-      PROMPTS_TO_CONSIDER: dream_persona
+      PROMPTS_TO_CONSIDER: dream_persona,dream_faq
       FLASK_APP: server
     context: .
     dockerfile: ./annotators/prompt_selector/Dockerfile

assistant_dists/document_based_qa/pipeline_conf.json

@@ -265,7 +265,7 @@
   "metadata": {
     "display_name": "ChatGPT-based Document QA",
     "author": "DeepPavlov",
-    "description": "This assistant uses the power of ChatGPT to answer you questions based on the document you provide.",
+    "description": "This assistant uses the power of ChatGPT to answer your questions based on the document you provide.",
     "version": "0.0.1",
     "date_created": "2023-01-10T02:00:00",
         "ram_usage": "9 GB",

assistant_dists/dream_persona_openai_prompted/dev.yml

@@ -10,12 +10,12 @@ services:
       - "./annotators/SentSeg:/src"
     ports:
       - 8011:8011
-  ranking-based-response-selector:
+  llm-based-response-selector:
     volumes:
-      - "./response_selectors/ranking_based_response_selector:/src"
+      - "./response_selectors/llm_based_response_selector:/src"
       - "./common:/src/common"
     ports:
-      - 8002:8002
+      - 8003:8003
   combined-classification:
     volumes:
       - "./common:/src/common"
@@ -52,5 +52,16 @@ services:
       - "./common:/src/common"
     ports:
       - 8162:8162
-
+  dff-dream-faq-prompted-skill:
+    volumes:
+      - "./skills/dff_template_prompted_skill:/src"
+      - "./common:/src/common"
+    ports:
+      - 8170:8170
+  openai-api-chatgpt-16k:
+    volumes:
+      - "./services/openai_api_lm:/src"
+      - "./common:/src/common"
+    ports:
+      - 8167:8167
 version: "3.7"

assistant_dists/dream_persona_openai_prompted/docker-compose.override.yml

@@ -2,9 +2,10 @@ services:
   agent:
     command: sh -c 'bin/wait && python -m deeppavlov_agent.run agent.pipeline_config=assistant_dists/dream_persona_openai_prompted/pipeline_conf.json'
     environment:
-      WAIT_HOSTS: "sentseg:8011, ranking-based-response-selector:8002, combined-classification:8087, 
+      WAIT_HOSTS: "sentseg:8011, llm-based-response-selector:8003, combined-classification:8087, 
                    sentence-ranker:8128, prompt-selector:8135, openai-api-chatgpt:8145,
-                   dff-dream-persona-chatgpt-prompted-skill:8137"
+                   dff-dream-persona-chatgpt-prompted-skill:8137, dff-dream-faq-prompted-skill:8170,
+          openai-api-chatgpt-16k:8167"
       WAIT_HOSTS_TIMEOUT: ${WAIT_TIMEOUT:-1000}
       HIGH_PRIORITY_INTENTS: 1
       RESTRICTION_FOR_SENSITIVE_CASE: 1
@@ -44,21 +45,22 @@ services:
         reservations:
           memory: 2G
 
-  ranking-based-response-selector:
-    env_file: [ .env ]
+  llm-based-response-selector:
+    env_file: [ .env,.env_secret ]
     build:
       args:
-        SERVICE_PORT: 8002
+        SERVICE_PORT: 8003
         SERVICE_NAME: response_selector
         LANGUAGE: EN
-        SENTENCE_RANKER_ANNOTATION_NAME: sentence_ranker
-        SENTENCE_RANKER_SERVICE_URL: http://sentence-ranker:8128/respond
-        SENTENCE_RANKER_TIMEOUT: 3
-        N_UTTERANCES_CONTEXT: 5
+        GENERATIVE_SERVICE_URL: http://openai-api-chatgpt:8145/respond
+        GENERATIVE_SERVICE_CONFIG: openai-chatgpt.json
+        GENERATIVE_TIMEOUT: 120
+        N_UTTERANCES_CONTEXT: 7
+        ENVVARS_TO_SEND: OPENAI_API_KEY,OPENAI_ORGANIZATION
         FILTER_TOXIC_OR_BADLISTED: 1
       context: .
-      dockerfile: ./response_selectors/ranking_based_response_selector/Dockerfile
-    command: flask run -h 0.0.0.0 -p 8002
+      dockerfile: ./response_selectors/llm_based_response_selector/Dockerfile
+    command: flask run -h 0.0.0.0 -p 8003
     environment:
       - FLASK_APP=server
     deploy:
@@ -75,7 +77,7 @@ services:
         SERVICE_PORT: 8135
         SERVICE_NAME: prompt_selector
         N_SENTENCES_TO_RETURN: 3
-        PROMPTS_TO_CONSIDER: dream_persona
+        PROMPTS_TO_CONSIDER: dream_persona,dream_faq
       context: .
       dockerfile: ./annotators/prompt_selector/Dockerfile
     command: flask run -h 0.0.0.0 -p 8135
@@ -164,4 +166,44 @@ services:
         reservations:
           memory: 128M
 
+  dff-dream-faq-prompted-skill:
+    env_file: [ .env,.env_secret ]
+    build:
+      args:
+        SERVICE_PORT: 8170
+        SERVICE_NAME: dff_dream_faq_prompted_skill
+        PROMPT_FILE: common/prompts/dream_faq.json
+        GENERATIVE_SERVICE_URL: http://openai-api-chatgpt-16k:8167/respond
+        GENERATIVE_SERVICE_CONFIG: openai-chatgpt.json
+        GENERATIVE_TIMEOUT: 120
+        N_UTTERANCES_CONTEXT: 7
+        ENVVARS_TO_SEND: OPENAI_API_KEY,OPENAI_ORGANIZATION
+      context: .
+      dockerfile: ./skills/dff_template_prompted_skill/Dockerfile
+    deploy:
+      resources:
+        limits:
+          memory: 128M
+        reservations:
+          memory: 128M
+
+  openai-api-chatgpt-16k:
+    env_file: [ .env ]
+    build:
+      args:
+        SERVICE_PORT: 8167
+        SERVICE_NAME: openai_api_chatgpt_16k
+        PRETRAINED_MODEL_NAME_OR_PATH: gpt-3.5-turbo-16k
+      context: .
+      dockerfile: ./services/openai_api_lm/Dockerfile
+    command: flask run -h 0.0.0.0 -p 8167
+    environment:
+      - FLASK_APP=server
+    deploy:
+      resources:
+        limits:
+          memory: 500M
+        reservations:
+          memory: 100M
+
 version: '3.7'

assistant_dists/dream_persona_openai_prompted/pipeline_conf.json

@@ -225,6 +225,27 @@
                     "component": "components/uYkoK0vRp4bbIg9akI1yw.yml",
                     "service": "skills/dummy_skill/service_configs/agent"
                 }
+            },
+            "dff_dream_faq_prompted_skill": {
+                "connector": {
+                    "protocol": "http",
+                    "timeout": 120.0,
+                    "url": "http://dff-dream-faq-prompted-skill:8170/respond"
+                },
+                "dialog_formatter": {
+                    "name": "state_formatters.dp_formatters:dff_prompted_skill_formatter",
+                    "skill_name": "dff_dream_faq_prompted_skill"
+                },
+                "response_formatter": "state_formatters.dp_formatters:skill_with_attributes_formatter_service",
+                "previous_services": [
+                    "skill_selectors"
+                ],
+                "state_manager_method": "add_hypothesis",
+                "is_enabled": true,
+                "source": {
+                    "component": "components/jFmKPqMJh0.yml",
+                    "service": "skills/dff_template_prompted_skill/service_configs/dff-dream-faq-prompted-skill"
+                }
             }
         },
         "candidate_annotators": {
@@ -269,8 +290,8 @@
             "response_selector": {
                 "connector": {
                     "protocol": "http",
-                    "timeout": 1.0,
-                    "url": "http://ranking-based-response-selector:8002/respond"
+                    "timeout": 120.0,
+                    "url": "http://llm-based-response-selector:8003/respond"
                 },
                 "dialog_formatter": "state_formatters.dp_formatters:cropped_dialog",
                 "response_formatter": "state_formatters.dp_formatters:base_response_selector_formatter_service",
@@ -281,7 +302,7 @@
                 "is_enabled": true,
                 "source": {
                     "component": "components/YJzc7NwGrLmKp6gfZJh7X1.yml",
-                    "service": "response_selectors/ranking_based_response_selector/service_configs/ranking-based-response-selector"
+                    "service": "response_selectors/llm_based_response_selector/service_configs/llm-based-response-selector"
                 }
             }
         }

response_selectors/llm_based_response_selector/server.py

@@ -120,7 +120,8 @@ def respond():
         hypotheses = [hyp for hyp in dialog["human_utterances"][-1]["hypotheses"]]
         if FILTER_TOXIC_OR_BADLISTED:
             hypotheses = filter_out_badlisted_or_toxic(hypotheses)
-
+        hypotheses_texts = "\n".join([f'{h["skill_name"]} (conf={h["confidence"]}): {h["text"]}' for h in hypotheses])
+        logger.info(f"Hypotheses: {hypotheses_texts}")
         dialog_context = [uttr["text"] for uttr in dialog["utterances"][-N_UTTERANCES_CONTEXT:]]
         selected_resp = select_response(
             dialog_context, hypotheses, dialog["human_utterances"][-1].get("attributes", {})

skill_selectors/description_based_skill_selector/connector.py

@@ -88,7 +88,7 @@ async def send(self, payload: Dict, callback: Callable):
                     skills_for_uttr.extend(prompted_skills)
                     logger.info("Adding all prompted skills as prompt selector did not select anything.")
 
-                if is_any_question_sentence_in_utterance(dialog["human_utterances"][-1]) and is_factoid:
+                if is_any_question_sentence_in_utterance(dialog["human_utterances"][-1]):
                     skills_for_uttr.append("dff_google_api_skill")
 
                 if is_factoid:

skills/dff_document_qa_llm_skill/README.md

@@ -0,0 +1,28 @@
+# LLM-based Q&A on Documents Skill
+
+## Description
+
+LLM-based Q&A on Documents Skill answers questions about long documents provided by the user. It passes on document chunks most relevant to the user's question alongside with an instruction and the dialog context as a prompt to ChatGPT.
+
+## Parameters
+
+```
+GENERATIVE_SERVICE_URL: LLM to utilize
+GENERATIVE_SERVICE_CONFIG: configuration file with generative parameters to utilize
+GENERATIVE_TIMEOUT: timeout for request to LLM
+N_UTTERANCES_CONTEXT: number of last utterances to consider as a dialog context
+ENVVARS_TO_SEND: API keys splitted by comma to get as env variables
+FILE_SERVER_TIMEOUT: timeout for request where files are stored
+DOCUMENT_PROMPT_FILE: file to get the instruction from (to insert into prompt guiding the Question Answering model)
+```
+
+## Dependencies
+
+- LLM service provided in `GENERATIVE_SERVICE_URL`
+- annotator Document Retriever (both endpoints)
+- API keys in environmental variables for key-required LLMs (OpenAI API, Anthropic API)
+
+
+## How to get OpenAI API key
+
+Go to OpenAI and find your Secret API key in your [user settings](https://platform.openai.com/account/api-keys).