Separate into steps and wrap filesystem access

.gitignore

@@ -0,0 +1,3 @@
+input
+memory
+TODO.md

ai.py

@@ -0,0 +1,38 @@
+
+import openai
+
+
+class AI:
+    def __init__(self, **kwargs):
+        self.kwargs = kwargs
+
+    def start(self, system, user):
+        messages = [
+                {"role": "system", "content": system},
+                {"role": "user", "content": user},
+            ]
+
+        return self.next(messages)
+
+    def fsystem(self, msg):
+        return {"role": "system", "content": msg}
+    
+    def fuser(self, msg):
+        return {"role": "user", "content": msg}
+
+    def next(self, messages, prompt=None):
+        if prompt:
+            messages = messages + [{"role": "user", "content": prompt}]
+
+        response = openai.ChatCompletion.create(
+            messages=messages,
+            **self.kwargs
+        )
+
+        chat = []
+        for chunk in response:
+            delta = chunk['choices'][0]['delta']
+            msg = delta.get('content', '')
+            print(msg, end="")
+            chat.append(msg)
+        return messages + [{"role": "assistant", "content": "".join(chat)}]

chat_to_files.py

@@ -21,13 +21,9 @@ def parse_chat(chat):# -> List[Tuple[str, str]]:
     return files
 
 
-def to_files(chat, path):
-    os.makedirs(path, exist_ok=True)
-
-    with open(os.path.join(path, 'all_output.txt'), "w") as f:
-        f.write(chat)
+def to_files(chat, workspace):
+    workspace['all_output.txt'] = chat
 
     files = parse_chat(chat)
     for file_name, file_content in files:
-        with open(os.path.join(path, file_name), "w") as f:
-            f.write(file_content)
+        workspace[file_name] = file_content

db.py

@@ -0,0 +1,28 @@
+
+from dataclasses import dataclass
+import os
+from pathlib import Path
+
+
+class DB:
+    def __init__(self, path):
+        self.path = Path(path).absolute()
+        os.makedirs(self.path, exist_ok=True)
+
+    def __getitem__(self, key):
+        with open(self.path / key) as f:
+            return f.read()
+
+    def __setitem__(self, key, val):
+        with open(self.path / key, 'w') as f:
+            f.write(val)
+
+
+# dataclass for all dbs:
+@dataclass
+class DBs:
+    memory: DB
+    logs: DB
+    identity: DB
+    input: DB
+    workspace: DB

identity/clarify

@@ -0,0 +1,5 @@
+You will improve instructions by reading:
+1. ORIGINAL INSTRUCTIONS
+2. CLARIFYING QUESTIONS AND ANSWERS
+
+As output you will give a new version of the original instruction but where the answers to the clarifying questions have been incorporated to make it completely clear.

identity/philosophy

@@ -0,0 +1,5 @@
+You almost always put different classes in different files
+
+Python toolbelt preferences:
+- pytest
+- dataclasses

identity/qa

@@ -0,0 +1,4 @@
+You will read instructions and NOT carry them out, only seek to CLARIFY them.
+You will carry out the steps:
+1. Write a list of super short bullets of areas that are unclear
+2. Ask for only one clarifying questions and wait for a reply

main.py

@@ -1,47 +1,45 @@
+import json
 import os
 import pathlib
 from typing import Optional
 import openai
 from chat_to_files import to_files
+from ai import AI
+from steps import STEPS
+from db import DB, DBs
 import typer
 
 
 app = typer.Typer()
+    
 
 
 @app.command()
 def chat(
-    engine: str = "gpt-4",
-    temperature: float = 0.0,
+    model: str = "gpt-4",
+    temperature: float = 0.1,
     max_tokens: int = 4096,
     n: int = 1,
     stream: bool = True,
-    system_prompt: str = typer.Argument("system", help="System prompt file"),
-    user_prompt: str = typer.Argument("user", help="User prompt file"),
-    code_to_file_path: Optional[str] = typer.Option(
+    input_path: str = typer.Argument(
+        None, help="input path"
+    ),
+    memory_path: str = typer.Argument(
+        None, help="memory path"
+    ),
+    workspace_path: Optional[str] = typer.Option(
         None, "--out", "-c", help="Code to file path"
     ),
 ):
 
-    # ensure file path corresponds to file in the same file as this script, using __file__
-    if system_prompt == "system":
-        # get folder of script
-        system_prompt = pathlib.Path(__file__).parent / system_prompt
-
-    if user_prompt == "user":
-        user_prompt = pathlib.Path(__file__).parent / user_prompt
+    if memory_path is None:
+        memory_path = pathlib.Path(__file__).parent / 'memory'
 
+    if input_path is None:
+        input_path = pathlib.Path(__file__).parent / 'input'
     
-    with open(system_prompt, "r") as f:
-        system_prompt = f.read()
-    with open(user_prompt, "r") as f:
-        user_prompt = f.read()
-    response = openai.ChatCompletion.create(
-        model=engine,
-        messages=[
-            {"role": "system", "content": system_prompt},
-            {"role": "user", "content": user_prompt},
-        ],
+    ai = AI(
+        model=model,
         temperature=temperature,
         max_tokens=max_tokens,
         n=n,
@@ -49,15 +47,19 @@ def chat(
         stop=None,
     )
 
-    chat = []
-    for chunk in response:
-        delta = chunk['choices'][0]['delta']
-        msg = delta.get('content', '')
-        print(msg, end="")
-        chat.append(msg)
+    dbs = DBs(
+        memory=DB(memory_path),
+        logs=DB(pathlib.Path(memory_path) / 'logs'),
+        input=DB(input_path),
+        workspace=DB(workspace_path),
+        identity=DB(pathlib.Path(__file__).parent / 'identity'),
+    )
 
-    if code_to_file_path is not None:
-        to_files("".join(chat), code_to_file_path)
+    run_prefix= workspace_path.split('/')[-1] + '_' if workspace_path is not None else ''
+
+    for step in STEPS:
+        messages = step(ai, dbs)
+        dbs.logs[run_prefix + step.__name__] = json.dumps(messages)
 
 
 if __name__ == "__main__":

steps.py

@@ -0,0 +1,51 @@
+from ast import List
+from dataclasses import dataclass
+from typing import Callable
+from ai import AI
+from chat_to_files import to_files
+
+from db import DBs
+from db import DB
+
+def setup_sys_prompt(dbs):
+    return dbs.identity['setup'] + '\nUseful to know:\n' + dbs.identity['philosophy']
+
+def setup(ai: AI, dbs: DBs):
+    messages = ai.start(setup_sys_prompt(dbs), dbs.input['main_prompt'])
+    to_files(messages[-1]['content'], dbs.workspace)
+    return messages
+
+def run_clarified(ai: AI, dbs: DBs):
+    messages = ai.start(setup_sys_prompt(dbs), dbs.input['main_prompt'])
+    to_files(messages[-1]['content'], DB(str(dbs.workspace.path)+'_clarified'))
+    return messages
+
+def clarify(ai: AI, dbs: DBs):
+    messages = [ai.fsystem(dbs.identity['qa'])]
+    user = dbs.input['main_prompt']
+    while True:
+        messages = ai.next(messages, user)
+        print()
+        user = input('Answer: ')
+        if not user or user == 'q':
+            break
+
+        user += '\nIs anything else unclear? Please ask more questions until instructions are sufficient to write the code.'
+
+    # TOOD: Stop using clarify prompt. Just append questions and answers to the main prompt.
+    prompt = dbs.identity['clarify']
+    messages = ai.next([ai.fsystem(prompt)] + messages[1:], prompt)
+    dbs.memory['clarified_prompt'] = messages[-1]['content']
+    return messages
+
+
+# STEPS: List[Callable[[AI, DBs], List]] = [
+STEPS=[
+    setup,
+    # clarify,
+    # run_clarified
+    # to_files,
+    # improve_files,
+    # run_tests,
+    # ImproveBasedOnHumanComments
+]

tests/test_db.py

@@ -0,0 +1,14 @@
+from ..db import DB
+
+
+def test_db():
+    # use /tmp for testing
+    db = DB('/tmp/test_db')
+    db['test'] = 'test'
+    assert db['test'] == 'test'
+    db['test'] = 'test2'
+    assert db['test'] == 'test2'
+    db['test2'] = 'test2'
+    assert db['test2'] == 'test2'
+    assert db['test'] == 'test2'
+    print('test_db passed')

user

@@ -1,75 +0,0 @@
-Instructions:
-We are writing a feature computation framework.
-
-It will mainly consist of FeatureBuilder classes.
-
-Each Feature Builder will have the methods:
-- get(key, context, cache):  To first check cache, and then go on to call dependencies to compute the feature. Returns value and hash of value.
-- dry_run(key, context):  To check that "type" of key will match input requirements of features
-- input_type(context):  That explains what dimensions key is applying to
-- output_type(context):  That explains what type the output is
-
-It will have the class attr:
-- deps:  list of FeatureBuilder classes
-
-Where it is unclear, please make assumptions and add a commend in the code about it
-
-Here is an example of Builders we want:
-
-ProductEmbeddingString:  takes product_id, queries the product_db and gets the title as a string
-ProductEmbedding: takes string and returns and embedding
-ProductEmbeddingDB: takes just `merchant` name, uses all product_ids and returns the blob that is a database of embeddings
-ProductEmbeddingSearcher: takes a string, constructs embeddingDB feature (note: all features are cached), embeds the string and searches the db
-LLMProductPrompt:  queries the ProductEmbeddingString, and formats a template that says "get recommendations for {title}"
-LLMSuggestions: Takes product_id, looks up prompts and gets list of suggestions of product descriptions
-LLMLogic: Takes the product_id, gets the LLM suggestions, embeds the suggestions, does a search, and returns a list of product_ids
-
-
-The LLMLogic is the logic_builder in a file such as this one:
-```
-def main(merchant, market):
-    cache = get_cache()
-    interaction_data_db = get_interaction_data_db()
-    product_db = get_product_db()
-    merchant_config = get_merchant_config(merchant)[merchant]
-
-    context = Context(
-        interaction_data_db=interaction_data_db,
-        product_db=product_db,
-        merchant_config=merchant_config,
-    )
-    
-    product_ids = cache(ProductIds.get)(
-        key=(merchant, market),
-        context=context,
-        cache=cache,
-    )
-
-    for logic_builder in merchant_config['logic_builders']:
-        for product_id in product_ids:
-            key = (merchant, market, product_id)
-            p2p_recs = cache(logic_builder.get)(key, cache, context)
-            redis.set(key, p2p_recs)
-```
-
-API to product_db:
-```
-    async def get_product_attribute_dimensions(
-        self,
-    ) -> dict[AttributeId, Dimension]:
-        return await self.repository.get_product_attribute_dimensions(self.merchant)
-
-    async def get_products(
-        self,
-        attribute_ids: set[AttributeId],
-        product_ids: set[ProductId] | None = None,
-    ) -> dict[ProductId, dict[AttributeId, dict[IngestionDimensionKey, Any]]]:
-        return await self.repository.get_products_dict(
-            self.merchant, attribute_ids, product_ids
-        )
-```
-
-(note, dimensions are not so important. They related to information that varies by: locale, warehouse, pricelist etc)
-
-
-Remember to read the Instructions carefully.