work through tutorial on function calling

bbbdb20d · Adrien Klose · 7ef0ba0a · bbbdb20d · bbbdb20d · bbbdb20d
Commit bbbdb20d authored 10 months ago by Adrien Klose
--- a/.gitignore
+++ b/.gitignore
+**/*.env
--- a/ideas_practice/code/LLM_structured_outputs_tutorial/functional_calling_test.ipynb
+++ b/ideas_practice/code/LLM_structured_outputs_tutorial/functional_calling_test.ipynb
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "True"
+      ]
+     },
+     "execution_count": 1,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "import os\n",
+    "import instructor\n",
+    "\n",
+    "from groq import Groq\n",
+    "from pydantic import BaseModel\n",
+    "import dotenv\n",
+    "from typing_extensions import Literal\n",
+    "from enum import Enum\n",
+    "dotenv.load_dotenv()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "client = Groq(\n",
+    "    api_key=os.environ.get(\"GROQ_API_KEY\"),\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# By default, the patch function will patch the ChatCompletion.create and ChatCompletion.create methods to support the response_model parameter\n",
+    "client = instructor.from_groq(client, mode=instructor.Mode.TOOLS)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Can use the response_model parameter using the BaseModel\n",
+    "#class QuestionType(Enum):\n",
+    "#    YesNo: Literal[\"Yes\",\"No\"]\n",
+    "#    Factoid: str\n",
+    "#    List: list[str]\n",
+    "#    Summary: str\n",
+    "\n",
+    "class YesNoModel(BaseModel):\n",
+    "    answer: Literal[\"Yes\",\"No\"]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "yesno: YesNoModel = client.chat.completions.create(\n",
+    "    model=\"mixtral-8x7b-32768\",\n",
+    "    response_model=YesNoModel,\n",
+    "    messages=[\n",
+    "        {\"role\": \"system\", \"content\": \"You are a helpful medical expert who answers with either yes or no to questions.\"},\n",
+    "        {\"role\": \"user\", \"content\": \"Is it safe to take isotretinoin during pregnancy?\"},\n",
+    "    ],\n",
+    ")\n",
+    "\n",
+    "assert isinstance(yesno, YesNoModel), \"Should be instance of UserExtract\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "answer='No'\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(yesno)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 20,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Not sure I understand the use of Field\n",
+    "from pydantic import Field\n",
+    "class NLSentenceModel(BaseModel):\n",
+    "    sentence: str = Field(...,description=\"Must be natural language sentences\")\n",
+    "# Could include a validator to check whether the result is a single sentence.\n",
+    "\n",
+    "sent: NLSentenceModel = client.chat.completions.create(\n",
+    "    model=\"mixtral-8x7b-32768\",\n",
+    "    response_model=NLSentenceModel,\n",
+    "    messages=[\n",
+    "        {\"role\": \"system\", \"content\": \"You are a helpful medical expert who reformulates triples of subjec, relationship and object into one sentence.\"},\n",
+    "        {\"role\": \"user\", \"content\": \"isotretinoin has_active_gen pregnancy\"},\n",
+    "    ],\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 21,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "sentence='Isotretinoin has an active genetic effect on pregnancy.'\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(sent)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.12"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}
+%% Cell type:code id: tags:
+
+``` python
+import os
+import instructor
+
+from groq import Groq
+from pydantic import BaseModel
+import dotenv
+from typing_extensions import Literal
+from enum import Enum
+dotenv.load_dotenv()
+```
+
+%% Output
+
+    True
+
+%% Cell type:code id: tags:
+
+``` python
+client = Groq(
+    api_key=os.environ.get("GROQ_API_KEY"),
+)
+```
+
+%% Cell type:code id: tags:
+
+``` python
+# By default, the patch function will patch the ChatCompletion.create and ChatCompletion.create methods to support the response_model parameter
+client = instructor.from_groq(client, mode=instructor.Mode.TOOLS)
+```
+
+%% Cell type:code id: tags:
+
+``` python
+# Can use the response_model parameter using the BaseModel
+#class QuestionType(Enum):
+#    YesNo: Literal["Yes","No"]
+#    Factoid: str
+#    List: list[str]
+#    Summary: str
+
+class YesNoModel(BaseModel):
+    answer: Literal["Yes","No"]
+```
+
+%% Cell type:code id: tags:
+
+``` python
+yesno: YesNoModel = client.chat.completions.create(
+    model="mixtral-8x7b-32768",
+    response_model=YesNoModel,
+    messages=[
+        {"role": "system", "content": "You are a helpful medical expert who answers with either yes or no to questions."},
+        {"role": "user", "content": "Is it safe to take isotretinoin during pregnancy?"},
+    ],
+)
+
+assert isinstance(yesno, YesNoModel), "Should be instance of UserExtract"
+```
+
+%% Cell type:code id: tags:
+
+``` python
+print(yesno)
+```
+
+%% Output
+
+    answer='No'
+
+%% Cell type:code id: tags:
+
+``` python
+# Not sure I understand the use of Field
+from pydantic import Field
+class NLSentenceModel(BaseModel):
+    sentence: str = Field(...,description="Must be natural language sentences")
+# Could include a validator to check whether the result is a single sentence.
+
+sent: NLSentenceModel = client.chat.completions.create(
+    model="mixtral-8x7b-32768",
+    response_model=NLSentenceModel,
+    messages=[
+        {"role": "system", "content": "You are a helpful medical expert who reformulates triples of subjec, relationship and object into one sentence."},
+        {"role": "user", "content": "isotretinoin has_active_gen pregnancy"},
+    ],
+)
+```
+
+%% Cell type:code id: tags:
+
+``` python
+print(sent)
+```
+
+%% Output
+
+    sentence='Isotretinoin has an active genetic effect on pregnancy.'
+
+%% Cell type:code id: tags:
+
+``` python
+```
--- a/ideas_practice/code/LLM_structured_outputs_tutorial/notes.txt
+++ b/ideas_practice/code/LLM_structured_outputs_tutorial/notes.txt
+Chapter 1 Asking LLM for structured data
+- Pydantic to give key and type security for dictionaries
+- Pydantic to structure LLM output
+- function calling for asking LLMs for json output; allows to
+  better use LLM that are fine tuned for json
+- other libraries include Marvin, Langchain, LlamaIndex
+
+Chapter 2 Prompting LLMs
+- Schemas to make Prompt engineering modular and reusable
+- Literals in schemes seem to outperform enums
+
+- we can extract arbitrary attributes too and limit the amount ???
+- can use this extract responses objects as they come with stream
+--> complexer requests improved upon their perceived latency
+
+- can use an index list with "good" prompt to extract relationships between entities
+- Making parts of the model optional counteracts hallucination!!!
+
+Chapter 3 RAG Applications
+- RAG = Retrieval Augmented Generation
+-> Retriever, Ranker, and Generator Component
+- embedded search = take query/prompt, get information from DB, put info and query into same prompt
+-> giving our own data allows model to answer even unseen questions/facts
+-> embedded search has limitations, model does not learn over time normally
+
+----> apply query understanding to match backend search capabilities
+- use query engine that breaks up query into multiple requests to backend
+- from the backend response we can generate summaries
+- can use wandb as a library for better documentation/logging
+
+- can use it to decompose questions into multiple and improve results
+
+Chapter 4 Validating LLM Outputs
+- use validators to generate feedback for LLM when its response did not confirm to our expectations
+- validators check all in parallel
+-> field function has lot of stuff already implemented
+- can make validation dynamique
+--> can use predefined like OpenAI moderation client
+-> can also make the llm check whether something is valid
+- can use validations to ensure the style of the respons
--- a/ideas_practice/code/bioasq_yesno_api_tests.ipynb
+++ b/ideas_practice/code/bioasq_yesno_api_tests.ipynb