Merge pull request #99 from fractalego/index-cache

Index cache

documentation/source/configuration.rst

@@ -12,8 +12,11 @@ A typical configuration file looks like this:
       "waking_up_sound": true,
       "deactivate_sound": true,
       "rules": "rules.yaml",
+      "index": "indices.yaml",
+      "cache_filename": "knowledge_cache",
+      "prompt_filename": "main.prompt",
       "functions": "functions.py",
-      "frontend_port": 8081,
+      "max_recursion": 2,
       "llm_model": {
         "model_host": "localhost",
         "model_port": 8080,
@@ -37,6 +40,7 @@ A typical configuration file looks like this:
     }
 
 
+
 These settings regulate the following:
 
     * "waking_up_word" is the name of the bot, used to wake up the system in the "run-audio" mode.
@@ -45,6 +49,12 @@ These settings regulate the following:
 
     * "rules" is the file containing the facts and rules that guide the chatbot. The default is "rules.yaml".
 
+    * "index" is the file containing the path to the files to index. The default is "indices.yaml".
+
+    * "cache_filename" is the file where the indexed knowledge is cached. The default is "knowledge_cache".
+
+    * "prompt_filename" is the file containing the main prompt for the chatbot. The default is "main.prompt".
+
     * "functions" is the file containing the functions that can be used in the rules. The default is "functions.py".
 
     * "frontend_port" is the port where the web frontend is running. The default is 8090.

documentation/source/index.rst

@@ -16,6 +16,7 @@ Welcome to WAFL's 0.0.90 documentation!
    configuration
    running_WAFL
    facts_and_rules
+   modify_the_prompt
    examples
    testcases
    actions

documentation/source/modify_the_prompt.rst

@@ -0,0 +1,23 @@
+Modify the original prompt
+==========================
+
+The prompt is stored in the file "main.prompt" in the project's root directory.
+The name of the file can be changed in the `config.json` file.
+The default is:
+
+
+.. code-block:: text
+    A user is chatting with a bot. The chat is happening through a web interface. The user is typing the messages and the bot is replying.
+
+    This is summary of the bot's knowledge:
+    {facts}
+
+    The rules that *must* be followed are:
+    {rules}
+
+    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.
+
+
+The variables `{facts}` and `{rules}` are replaced by the actual facts and rules when the prompt is generated.

requirements.txt

@@ -17,4 +17,7 @@ sphinx-rtd-theme==1.2.0
 bluepy==1.3.0
 einops==0.6.1
 g2p-en==2.1.0
-pyyaml==6.0.1
+pyyaml==6.0.1
+joblib==1.4.2
+pymupdf==1.24.7
+

setup.py

@@ -22,6 +22,7 @@
         "wafl.connectors.remote",
         "wafl.events",
         "wafl.extractors",
+        "wafl.handlers",
         "wafl.inference",
         "wafl.interface",
         "wafl.knowledge",
@@ -61,6 +62,8 @@
         "einops==0.6.1",
         "g2p-en==2.1.0",
         "pyyaml==6.0.1",
+        "joblib==1.4.2",
+        "pymupdf==1.24.7",
     ],
     classifiers=[
         "License :: OSI Approved :: MIT License",

tests/config.json

@@ -3,6 +3,9 @@
   "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": {

tests/indices.yaml

@@ -0,0 +1,2 @@
+paths:
+- files_to_index/

tests/main.prompt

@@ -0,0 +1,11 @@
+A user is chatting with a bot. The chat is happening through a web interface. The user is typing the messages and the bot is replying.
+
+This is summary of the bot's knowledge:
+{facts}
+
+The rules that *must* be followed are:
+{rules}
+
+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.

tests/test_indexing.py

@@ -0,0 +1,49 @@
+import asyncio
+import os
+import yaml
+
+from unittest import TestCase
+
+from wafl.config import Configuration
+from wafl.dataclasses.dataclasses import Query
+from wafl.knowledge.indexing_implementation import add_to_index, load_knowledge
+
+_path = os.path.dirname(__file__)
+
+
+class TestIndexing(TestCase):
+    def test__path_can_be_added_to_index(self):
+        data = _load_index()
+        prior_count = len(data["paths"])
+        add_to_index("files_to_index2")
+
+        data = _load_index()
+        current_count = len(data["paths"])
+        self.assertEqual(current_count, prior_count + 1)
+
+        data["paths"].remove("files_to_index2")
+        with open("indices.yaml", "w") as file:
+            file.write(yaml.dump(data))
+
+    def test__indexed_files_can_be_retrieved(self):
+        config = Configuration.load_local_config()
+        knowledge = asyncio.run(load_knowledge(config))
+        results = asyncio.run(
+            knowledge.ask_for_facts(Query.create_from_text("How do I start WAFL"))
+        )
+        expected = "WAFL"
+        self.assertIn(expected, results[0].text)
+
+    def test__pdf_can_be_read(self):
+        config = Configuration.load_local_config()
+        knowledge = asyncio.run(load_knowledge(config))
+        results = asyncio.run(
+            knowledge.ask_for_facts(Query.create_from_text("What color is the sky?"))
+        )
+        expected = "green"
+        self.assertIn(expected, results[0].text)
+
+
+def _load_index():
+    with open("indices.yaml", "r") as file:
+        return yaml.safe_load(file.read())

tests/test_rules.py

@@ -39,8 +39,8 @@ def test__rules_can_be_triggered(self):
             interface=interface,
         )
         asyncio.run(conversation_events.process_next())
-        expected = "The horse is tall"
-        self.assertIn(expected, interface.get_utterances_list()[-1])
+        expected = "the horse is tall"
+        self.assertIn(expected, interface.get_utterances_list()[-1].lower())
 
     def test__rules_are_not_always_triggered(self):
         interface = DummyInterface(

todo.txt

@@ -1,5 +1,39 @@
+* interruptible speech
+* dependabot!!!
+* use poetry
+
+PharazonE
+* upload to hetzner and make it work for some retrieval tasks
+* develop more rules + use-cases for voice and other
+
+
+/* add control over which llm to use from the frontend
+/  - add list of models in the backend
+
+/* add quantization of llm to wafl_llm config
+/* write docs about it on wafl
+
+/* add option so use llama.cpp from wafl_llm
+/* add option to have None as a model setting in wafl_llm
+
+/* add pdf to indexing
+* add json to indexing
+/* add metadata to indexing items
+
+
+/make backend it run with ollama as well (no too slow)
+
+
+/None of the knowledge is loaded from the web interface. Why?
+/- you have just changed the load_knowledge function to make it async.
+
+
 wafl:
-- create indices
+/- create indices
+- allow files/folders to be indexed (modify rules.yaml and then re-index)
+- add keywords in retrieval from tfidf
+- silence output when someone speak
+- multiple models in wafl-llm, with selection from frontend
 
 training:
 - retrain phi3

wafl/answerer/answerer_implementation.py

@@ -4,7 +4,7 @@
 from typing import List, Tuple
 
 from wafl.exceptions import CloseConversation
-from wafl.facts import Fact
+from wafl.dataclasses.facts import Fact
 from wafl.interface.conversation import Conversation, Utterance
 
 
@@ -116,7 +116,15 @@ async def _run_code(to_execute: str, module, functions) -> str:
 def get_text_from_facts_and_thresholds(
     facts_and_thresholds: List[Tuple[Fact, float]], memory: str
 ) -> List[str]:
-    return [item[0].text for item in facts_and_thresholds if item[0].text not in memory]
+    text_list = []
+    for item in facts_and_thresholds:
+        if item[0].text not in memory:
+            text = item[0].text
+            if item[0].metadata:
+                text = f"Metadata for the following text: {str(item[0].metadata)}" + "\n" + text
+            text_list.append(text)
+
+    return text_list
 
 
 def add_dummy_utterances_to_continue_generation(

wafl/answerer/dialogue_answerer.py

@@ -2,7 +2,6 @@
 from inspect import getmembers, isfunction
 from typing import List, Tuple
 from wafl.answerer.answerer_implementation import (
-    is_executable,
     substitute_memory_in_answer_and_get_memories_if_present,
     create_one_liner,
     get_text_from_facts_and_thresholds,
@@ -12,17 +11,16 @@
 )
 from wafl.answerer.base_answerer import BaseAnswerer
 from wafl.answerer.rule_maker import RuleMaker
-from wafl.connectors.clients.llm_chitchat_answer_client import LLMChitChatAnswerClient
-from wafl.extractors.dataclasses import Query, Answer
-from wafl.interface.conversation import Conversation, Utterance
+from wafl.connectors.clients.llm_chat_client import LLMChatClient
+from wafl.dataclasses.dataclasses import Query, Answer
+from wafl.interface.conversation import Conversation
 from wafl.simple_text_processing.questions import is_question
 
 
 class DialogueAnswerer(BaseAnswerer):
     def __init__(self, config, knowledge, interface, code_path, logger):
         self._threshold_for_facts = 0.85
-        self._delete_current_rule = "[delete_rule]"
-        self._client = LLMChitChatAnswerClient(config)
+        self._client = LLMChatClient(config)
         self._knowledge = knowledge
         self._logger = logger
         self._interface = interface
@@ -38,7 +36,6 @@ def __init__(self, config, knowledge, interface, code_path, logger):
             config,
             interface,
             max_num_rules=1,
-            delete_current_rule=self._delete_current_rule,
         )
 
     async def answer(self, query_text: str) -> Answer:
@@ -51,10 +48,6 @@ async def answer(self, query_text: str) -> Answer:
         rules_text = await self._get_relevant_rules(conversation)
         if not conversation:
             conversation = create_one_liner(query_text)
-        last_bot_utterances = conversation.get_last_speaker_utterances("bot", 3)
-        last_user_utterance = conversation.get_last_speaker_utterances("user", 1)
-        if not last_user_utterance:
-            last_user_utterance = query_text
         conversational_timestamp = len(conversation)
         facts = await self._get_relevant_facts(
             query,
@@ -73,18 +66,9 @@ async def answer(self, query_text: str) -> Answer:
             answer_text, memories = await self._apply_substitutions(
                 original_answer_text
             )
-            if answer_text in last_bot_utterances and not is_executable(
-                original_answer_text
-            ):
-                conversation = create_one_liner(last_user_utterance[-1])
-                continue
-
-            if self._delete_current_rule in answer_text:
-                self._prior_rules = []
-                final_answer_text += answer_text
-                break
 
             final_answer_text += answer_text
+
             if not memories:
                 break
 
@@ -130,7 +114,7 @@ async def _get_relevant_rules(self, conversation: Conversation) -> List[str]:
         for rule in rules:
             if rule not in self._prior_rules:
                 self._prior_rules.insert(0, rule)
-        self._prior_rules = self._prior_rules[:self._max_num_past_utterances_for_rules]
+        self._prior_rules = self._prior_rules[: self._max_num_past_utterances_for_rules]
         return self._prior_rules
 
     def _init_python_module(self, module_name):

wafl/answerer/rule_maker.py

@@ -1,7 +1,7 @@
 from typing import List
 
-from wafl.extractors.dataclasses import Query
-from wafl.rules import Rule
+from wafl.dataclasses.dataclasses import Query
+from wafl.dataclasses.rules import Rule
 
 
 class RuleMaker:
@@ -11,14 +11,12 @@ def __init__(
         config: "BaseConfig",
         interface: "BaseInterface",
         max_num_rules: int,
-        delete_current_rule: str,
         max_recursion: int = 3,
     ):
         self._knowledge = knowledge
         self._config = config
         self._interface = interface
         self._max_num_rules = max_num_rules
-        self._delete_current_rule = delete_current_rule
         if not config.get_value("max_recursion"):
             self._max_indentation = max_recursion
         else:
@@ -29,9 +27,7 @@ async def create_from_query(self, conversation: "Conversation") -> List[str]:
         rules_texts = []
         for rule in rules:
             rules_text = rule.get_string_using_template(
-                "- {effect}:\n{clauses}\n"
-                + rule.indent_str
-                + f'- After you completed all the steps output "{self._delete_current_rule}".\n'
+                "- {effect}:\n{clauses}\n" + rule.indent_str
             )
             rules_texts.append(rules_text)
             await self._interface.add_fact(f"The bot remembers the rule:\n{rules_text}")