Development

documentation/source/configuration.rst

@@ -7,37 +7,33 @@ A typical configuration file looks like this:
 
 .. code-block:: text
 
-    {
-      "waking_up_word": "computer",
-      "waking_up_sound": true,
-      "deactivate_sound": true,
-      "rules": "rules.yaml",
-      "index": "indices.yaml",
-      "cache_filename": "knowledge_cache",
-      "prompt_filename": "main.prompt",
-      "functions": "functions.py",
-      "max_recursion": 2,
-      "llm_model": {
-        "model_host": "localhost",
-        "model_port": 8080,
-        "temperature": 0.4
-      },
-      "listener_model": {
-        "model_host": "localhost",
-        "model_port": 8080,
-        "listener_hotword_logp": -8,
-        "listener_volume_threshold": 0.6,
-        "listener_silence_timeout": 0.7
-      },
-      "speaker_model": {
-        "model_host": "localhost",
-        "model_port": 8080
-      },
-      "text_embedding_model": {
-        "model_host": "localhost",
-        "model_port": 8080
-      }
-    }
+{
+  "waking_up_word": "computer",
+  "waking_up_sound": true,
+  "deactivate_sound": true,
+  "rules": "rules.yaml",
+  "index": "indices.yaml",
+  "cache_filename": "knowledge_cache",
+  "prompt_filename": "main.prompt",
+  "functions": "functions.py",
+  "max_recursion": 2,
+  "frontend_port": 8090,
+  "backend": {
+    "host": "localhost",
+    "port": 8080,
+    "token": "secret"
+  },
+  "generation_config": {
+    "temperature": 0.4
+  },
+  "listener_model": {
+    "listener_hotword_logp": -8,
+    "listener_volume_threshold": 0.6,
+    "listener_silence_timeout": 0.7,
+    "interruptible": true
+  }
+}
+
 
 
 
@@ -59,20 +55,10 @@ These settings regulate the following:
 
     * "frontend_port" is the port where the web frontend is running. The default is 8090.
 
-    * "llm_model" is the configuration to connect to wafl-llm in the backend. The default url is "localhost:8080". The "temperature" parameter is used to set the temperature for the LLM model. The default is 0.4.
-
-    * "listener_model" is the configuration to connect to the listener model in the backend. The default is "localhost:8080".
-
-       - The listener model is used to detect the wake-up word.
-         The similarity threshold for the detection can be set with the "listener_hotword_logp" parameter.
-
-       - The "listener_volume_threshold" parameter is used to set the volume threshold for any conversation.
-         Any word uttered with a volume below this threshold is ignored.
-
-       - The "listener_silence_timeout" parameter is used to set the silence timeout for any conversation.
-         If no word is uttered for a time longer than this timeout, the conversation is considered finished.
-
-    * "speaker_model" is the configuration to connect to the speaker model in the backend. The default is "localhost:8080".
+    * "backend" is the configuration related to the backend. The default is "localhost:8080".
 
-    * "text_embedding_model" is the configuration to connect to the text embedding model in the backend. The default is "localhost:8080".
+    * "generation_config" is the configuration related to the generation of the response. The default is "temperature: 0.4".
 
+    * "listener_model" is the configuration related to the listener model.
+      These items determine the thresholds for hotword detection, volume threshold, silence timeout, and whether the listener is interruptible.
+      The default is "listener_hotword_logp: -8", "listener_volume_threshold: 0.6", "listener_silence_timeout: 0.7", "interruptible: true".

requirements.txt

@@ -1,7 +1,6 @@
 flask[async]==3.0.3
 flask-cors==4.0.1
 nltk==3.8.1
-gensim==4.3.3
 sklearn==0.0
 python-Levenshtein==0.25.1
 fuzzywuzzy==0.18.0

setup.py

@@ -20,9 +20,8 @@
         "wafl.connectors.clients",
         "wafl.connectors.factories",
         "wafl.connectors.remote",
-        "wafl.dataclasses",
+        "wafl.data_objects",
         "wafl.events",
-        "wafl.extractors",
         "wafl.handlers",
         "wafl.inference",
         "wafl.interface",
@@ -49,7 +48,6 @@
         "flask[async]==3.0.3",
         "flask-cors==4.0.1",
         "nltk==3.8.1",
-        "gensim==4.3.3",
         "sklearn==0.0",
         "python-Levenshtein==0.25.1",
         "fuzzywuzzy==0.18.0",

tests/config.json

@@ -8,24 +8,19 @@
   "prompt_filename": "main.prompt",
   "functions": "functions.py",
   "max_recursion": 2,
-  "llm_model": {
-    "model_host": "localhost",
-    "model_port": 8080,
+  "frontend_port": 8090,
+  "backend": {
+    "host": "aragorn",
+    "port": 8080,
+    "token": "secret"
+  },
+  "generation_config": {
     "temperature": 0.4
   },
   "listener_model": {
-    "model_host": "localhost",
-    "model_port": 8080,
     "listener_hotword_logp": -8,
     "listener_volume_threshold": 0.6,
-    "listener_silence_timeout": 0.7
-  },
-  "speaker_model": {
-    "model_host": "localhost",
-    "model_port": 8080
-  },
-  "text_embedding_model": {
-    "model_host": "localhost",
-    "model_port": 8080
+    "listener_silence_timeout": 0.7,
+    "interruptible": true
   }
 }

tests/main.prompt

@@ -8,4 +8,6 @@ The rules that *must* be followed are:
 
 Create a plausible dialogue based on the aforementioned summary and rules.
 Do not repeat yourself. Be friendly but not too servile.
-Follow the rules if present and they apply to the dialogue. Do not improvise if rules are present.
+Follow the rules if present and they apply to the dialogue. Do not improvise if rules are present.
+The user query might be incomplete or ambiguous or ungrammatical. The bot *must* ask for clarification if needed.
+The bot only answers if the query is clear and unambiguous.

tests/test_connection.py

@@ -12,7 +12,7 @@
 class TestConnection(TestCase):
     def test__connection_to_generative_model_can_generate_text(self):
         config = Configuration.load_local_config()
-        connector = RemoteLLMConnector(config.get_value("llm_model"))
+        connector = RemoteLLMConnector(config)
         prediction = asyncio.run(
             connector.predict(
                 PromptCreator.create_from_one_instruction(
@@ -25,7 +25,7 @@ def test__connection_to_generative_model_can_generate_text(self):
 
     def test__connection_to_generative_model_can_generate_text_within_tags(self):
         config = Configuration.load_local_config()
-        connector = RemoteLLMConnector(config.get_value("llm_model"))
+        connector = RemoteLLMConnector(config)
         connector._num_prediction_tokens = 200
         text = 'Generate a full paragraph based on this chapter title " The First Contact". The theme of the paragraph is space opera. Include the characters "Alberto" and "Maria". Write at least three sentences.'
         prompt = f"""
@@ -43,7 +43,7 @@ def test__connection_to_generative_model_can_generate_text_within_tags(self):
 
     def test__connection_to_generative_model_can_generate_a_python_list(self):
         config = Configuration.load_local_config()
-        connector = RemoteLLMConnector(config.get_value("llm_model"))
+        connector = RemoteLLMConnector(config)
         connector._num_prediction_tokens = 200
         prompt = "Generate a Python list of 4 chapters names for a space opera book. The output needs to be a python list of strings: "
         prediction = asyncio.run(

tests/test_entailer.py

@@ -0,0 +1,34 @@
+import asyncio
+import os
+
+from unittest import TestCase
+from wafl.config import Configuration
+from wafl.connectors.remote.remote_entailer_connector import RemoteEntailerConnector
+from wafl.connectors.clients.entailer_client import EntailerClient
+
+_path = os.path.dirname(__file__)
+
+
+class TestConnection(TestCase):
+    def test__entailer_connector(self):
+        config = Configuration.load_local_config()
+        connector = RemoteEntailerConnector(config)
+        prediction = asyncio.run(
+            connector.predict(
+                "The first contact is a romance novel set in the middle ages.",
+                "The first contact is a science fiction novel about the first contact between humans and aliens.",
+            )
+        )
+        assert prediction["score"] < 0.5
+
+    def test__entailment_client(self):
+
+        config = Configuration.load_local_config()
+        client = EntailerClient(config)
+        prediction = asyncio.run(
+            client.get_entailment_score(
+                "The first contact is a romance novel set in the middle ages.",
+                "The first contact is a science fiction novel about the first contact between humans and aliens.",
+            )
+        )
+        assert prediction < 0.5

tests/test_indexing.py

@@ -5,7 +5,7 @@
 from unittest import TestCase
 
 from wafl.config import Configuration
-from wafl.dataclasses.dataclasses import Query
+from wafl.data_objects.dataclasses import Query
 from wafl.knowledge.indexing_implementation import add_to_index, load_knowledge
 
 _path = os.path.dirname(__file__)

tests/test_speaker.py

@@ -4,7 +4,7 @@
 from unittest import TestCase
 
 from wafl.config import Configuration
-from wafl.speaker.fairseq_speaker import FairSeqSpeaker
+from wafl.speaker.tts_speaker import TTSSpeaker
 from wafl.speaker.soundfile_speaker import SoundFileSpeaker
 
 _wafl_greetings = """
@@ -17,24 +17,30 @@
 class TestSpeaker(TestCase):
     def test_voice(self):
         config = Configuration.load_local_config()
-        speaker = FairSeqSpeaker(config)
+        speaker = TTSSpeaker(config)
         text = "Hello world"
         asyncio.run(speaker.speak(text))
 
     def test_long_text(self):
         config = Configuration.load_local_config()
-        speaker = FairSeqSpeaker(config)
+        speaker = TTSSpeaker(config)
         text = (
             "Shall I compare you to a summer's day? Thou art more lovely and temperate."
         )
         asyncio.run(speaker.speak(text))
 
-    def test_number_pronunciation(self):
+    def test_number_pronunciation1(self):
         config = Configuration.load_local_config()
-        speaker = FairSeqSpeaker(config)
+        speaker = TTSSpeaker(config)
         text = "The time is 54 past 8"
         asyncio.run(speaker.speak(text))
 
+    def test_number_pronunciation2(self):
+        config = Configuration.load_local_config()
+        speaker = TTSSpeaker(config)
+        text = "The time is 8 54"
+        asyncio.run(speaker.speak(text))
+
     def test_on_sound(self):
         speaker = SoundFileSpeaker()
         speaker.speak(os.path.join(_path, "../wafl/sounds/activation.wav"))

tests/test_voice.py

@@ -15,15 +15,15 @@
 
 rules:
   - the user's name is Jane:
-    - write "I hear you"
+    - reply with "I hear you" and nothing else
 """.strip()
 
 _path = os.path.dirname(__file__)
 
 
 class TestVoice(TestCase):
     def test__activation(self):
-        interface = DummyInterface(to_utter=["computer", "my name is Jane"])
+        interface = DummyInterface(to_utter=["computer my name is Jane"])
         config = Configuration.load_local_config()
         config.set_value("rules", _wafl_example)
         conversation_events = ConversationEvents(config=config, interface=interface)

todo.txt

@@ -1,4 +1,42 @@
-* why do I need to re-initialise the retrievers after unpickling the knowledge?
+* user prior text and response from failed substitutions between [] instead of just iteration number (line 89, dialogue_answerer.py)
+* remove dead code
+* re-fine-tune phi to get better performance
+
+* delete rules and memory from discourse_answerer
+
+* This is wrong - from wafl_ll
+    <|end|><|assistant|><|user|> Hi!<|end|><|assistant|>
+   The user is sandwiched between the assistant. It should be:
+    <|end|><|assistant|> Hi!<|end|><|user|>
+
+/* make interruptible speech optional
+
+
+* use entailment score to flag a rule for execution before the answer.
+* get all model list from wafl_llm backend. Only specify the connection port and host in wafl
+
+* the answer from the indexed files should be directed from a rule.
+   - facts and rules should live at the highest level of the retrieval
+
+
+/* apply entailer to rule retrieval:
+/  if more than one rule is retrieved, then the one
+/  that is entailed by the query should be chosen
+
+
+
+/* Add tqdm to indexing.
+/* Make it index when wafl start first, not at the first use/login
+
+/* The prior items with timestamps might not be necessary.
+/  - Just implement a queue with a fixed size
+
+* add entailer to wafl_llm
+
+
+/* why do I need to re-initialise the retrievers after unpickling the knowledge?
+   - maybe you should save the retrievers in the knowledge object separately?
+   - It was gensim that was not serializable. Took it out
 
 /* knowledge cache does not cache the rules or facts