bulk-chain 0.25.3__tar.gz → 1.1.0__tar.gz

{bulk_chain-0.25.3 → bulk_chain-1.1.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: bulk_chain
-Version: 0.25.3
+Version: 1.1.0
 Summary: A lightweight, no-strings-attached Chain-of-Thought framework for your LLM, ensuring reliable results for bulk input requests.
 Home-page: https://github.com/nicolay-r/bulk-chain
 Author: Nicolay Rusnachenko
@@ -15,10 +15,8 @@ Classifier: Topic :: Text Processing :: Linguistic
 Requires-Python: >=3.6
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: tqdm
-Requires-Dist: source-iter==0.24.3
 
-# bulk-chain 0.25.3
+# bulk-chain 1.1.0
 ![](https://img.shields.io/badge/Python-3.9-brightgreen.svg)
 [![](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/nicolay-r/bulk-chain/blob/master/bulk_chain_tutorial.ipynb)
 [![twitter](https://img.shields.io/twitter/url/https/shields.io.svg?style=social)](https://x.com/nicolayr_/status/1847969224636961033)
@@ -31,7 +29,7 @@ Requires-Dist: source-iter==0.24.3
 <p align="center">
   <a href="https://github.com/nicolay-r/nlp-thirdgate?tab=readme-ov-file#llm"><b>Third-party providers hosting</b>↗️</a>
   <br>
-  <a href="https://github.com/nicolay-r/bulk-chain/blob/master/README.md#demo-mode">👉<b>demo</b>👈</a>
+  <a href="https://github.com/nicolay-r/bulk-chain-shell">👉<b>demo</b>👈</a>
 </p>
 
 A no-strings-attached **framework**  for your LLM that allows applying Chain-of-Thought-alike [prompt `schema`](#chain-of-thought-schema) towards a massive textual collections using custom **[third-party providers ↗️](https://github.com/nicolay-r/nlp-thirdgate?tab=readme-ov-file#llm)**.
@@ -39,11 +37,7 @@ A no-strings-attached **framework**  for your LLM that allows applying Chain-of-
 ### Main Features
 * ✅ **No-strings**: you're free to LLM dependencies and flexible `venv` customization.
 * ✅ **Support schemas descriptions** for Chain-of-Thought concept.
-* ✅ **Provides iterator over infinite amount of input contexts** served in `CSV`/`JSONL`.
-
-### Extra Features
-* ✅ **Progress caching [for remote LLMs]**: withstanding exception during LLM calls by using `sqlite3` engine for caching LLM answers;
-
+* ✅ **Provides iterator over infinite amount of input contexts**
 
 # Installation
 
@@ -83,60 +77,37 @@ Below, is an example on how to declare your own schema:
 
 # Usage
 
-Preliminary steps:
-
-1. Define your [schema](#chain-of-thought-schema) ([Example for Sentiment Analysis](/ext/schema/thor_cot_schema.json)))
-2. Wrap or pick **LLM model** from the [<b>Third-party providers hosting</b>↗️](https://github.com/nicolay-r/nlp-thirdgate?tab=readme-ov-file#llm).
-
-## Shell
-
-### Demo Mode
-
-**demo mode** to interact with LLM via command line with LLM output streaming support. 
-The video below illustrates an example of application for sentiment analysis on author opinion extraction towards mentioned object in text.
-
-Quck start with launching demo:
-1. ⬇️ Download [replicate](https://replicate.com/) provider for `bulk-chain`:
-2. 📜 Setup your reasoning `thor_cot_schema.json` according to the [following example ↗️](test/schema/thor_cot_schema.json)
-3. 🚀 Launch `demo.py` as follows:
-```bash
-python3 -m bulk_chain.demo \
-    --schema "test/schema/thor_cot_schema.json" \
-    --adapter "dynamic:replicate_104.py:Replicate" \
-    %%m \
-    --model_name "meta/meta-llama-3-70b-instruct" \
-    --api_token "<REPLICATE-API-TOKEN>" \
-    --stream
-```
-
-📺 This video showcase application of the [↗️ Sentiment Analysis Schema](https://github.com/nicolay-r/bulk-chain/blob/master/test/schema/thor_cot_schema.json) towards [LLaMA-3-70B-Instruct](https://replicate.com/meta/meta-llama-3-70b-instruct) hosted by Replicate for reasoning over submitted texts
-![sa-bulk-chain-cot-final](https://github.com/user-attachments/assets/0cc8fdcb-6ddb-44a3-8f05-d76250ae6423)
+## 🤖 Prepare 
 
+1. [schema](#chain-of-thought-schema)
+    * [Example for Sentiment Analysis](test/schema/thor_cot_schema.json)
+2. **LLM model** from the [<b>Third-party providers hosting</b>↗️](https://github.com/nicolay-r/nlp-thirdgate?tab=readme-ov-file#llm).
+3. Data (iter of dictionaries)
 
-### Inference Mode
+## 🚀 Launch
 
-> **NOTE:** You have to install `source-iter` and `tqdm` packages that actual [dependencies](dependencies.txt) of this project
+> **API**: For more details see the [**related Wiki page**](https://github.com/nicolay-r/bulk-chain/wiki)
 
-1. ⬇️ Download [replicate](https://replicate.com/) provider for `bulk-chain`:
-```bash
-wget https://raw.githubusercontent.com/nicolay-r/nlp-thirdgate/refs/heads/master/llm/replicate_104.py
-```
-2. 📜 Setup your reasoning `schema.json` according to the [following example ↗️](test/schema/default.json)
-3. 🚀 Launch inference using `DeepSeek-R1`:
-```bash
-python3 -m bulk_chain.infer \
-    --src "<PATH-TO-YOUR-CSV-or-JSONL>" \
-    --schema "test/schema/default.json" \
-    --adapter "replicate_104.py:Replicate" \
-    %%m \
-    --model_name "deepseek-ai/deepseek-r1" \
-    --api_token "<REPLICATE-API-TOKEN>"
+```python
+from bulk_chain.core.utils import dynamic_init
+from bulk_chain.api import iter_content
+
+content_it = iter_content(
+    # 1. Your schema.              
+    schema="YOUR_SCHEMA.json",
+    # 2. Your third-party model implementation.
+    llm=dynamic_init(class_filepath="replicate_104.py", class_name="Replicate")(api_token="<API-KEY>"),
+    # 3. Customize your inference and result providing modes: 
+    infer_mode="batch_async", 
+    return_mode="batch",
+    # 4. Your iterator of dictionaries
+    input_dicts_it=YOUR_DATA_IT,
+)
+    
+for content in content_it:
+    # Handle your LLM responses here ...
 ```
 
-## API
-
-Please take a look at the [**related Wiki page**](https://github.com/nicolay-r/bulk-chain/wiki)
-
 
 # Embed your LLM
 

bulk_chain-1.1.0/README.md

@@ -0,0 +1,100 @@
+# bulk-chain 1.1.0
+![](https://img.shields.io/badge/Python-3.9-brightgreen.svg)
+[![](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/nicolay-r/bulk-chain/blob/master/bulk_chain_tutorial.ipynb)
+[![twitter](https://img.shields.io/twitter/url/https/shields.io.svg?style=social)](https://x.com/nicolayr_/status/1847969224636961033)
+[![PyPI downloads](https://img.shields.io/pypi/dm/bulk-chain.svg)](https://pypistats.org/packages/bulk-chain)
+
+<p align="center">
+    <img src="logo.png"/>
+</p>
+
+<p align="center">
+  <a href="https://github.com/nicolay-r/nlp-thirdgate?tab=readme-ov-file#llm"><b>Third-party providers hosting</b>↗️</a>
+  <br>
+  <a href="https://github.com/nicolay-r/bulk-chain-shell">👉<b>demo</b>👈</a>
+</p>
+
+A no-strings-attached **framework**  for your LLM that allows applying Chain-of-Thought-alike [prompt `schema`](#chain-of-thought-schema) towards a massive textual collections using custom **[third-party providers ↗️](https://github.com/nicolay-r/nlp-thirdgate?tab=readme-ov-file#llm)**.
+
+### Main Features
+* ✅ **No-strings**: you're free to LLM dependencies and flexible `venv` customization.
+* ✅ **Support schemas descriptions** for Chain-of-Thought concept.
+* ✅ **Provides iterator over infinite amount of input contexts**
+
+# Installation
+
+From PyPI: 
+
+```bash
+pip install --no-deps bulk-chain
+```
+
+or latest version from here:
+
+```bash
+pip install git+https://github.com/nicolay-r/bulk-chain@master
+```
+
+## Chain-of-Thought Schema
+
+To declare Chain-of-Though (CoT) schema, this project exploits `JSON` format.
+This format adopts `name` field for declaring a name and `schema` is a list of CoT instructions for the Large Language Model.
+
+Each step represents a dictionary with `prompt` and `out` keys that corresponds to the input prompt and output variable name respectively.
+All the variable names are expected to be mentioned in `{}`.
+
+Below, is an example on how to declare your own schema:
+
+```python
+{
+"name": "schema-name",
+"schema": [
+    {"prompt": "Given the question '{text}', let's think step-by-step.", 
+     "out": "steps"},
+    {"prompt": "For the question '{text}' the reasoining steps are '{steps}'. what would be an answer?", 
+     "out":  "answer"},
+]
+}
+```
+
+# Usage
+
+## 🤖 Prepare 
+
+1. [schema](#chain-of-thought-schema)
+    * [Example for Sentiment Analysis](test/schema/thor_cot_schema.json)
+2. **LLM model** from the [<b>Third-party providers hosting</b>↗️](https://github.com/nicolay-r/nlp-thirdgate?tab=readme-ov-file#llm).
+3. Data (iter of dictionaries)
+
+## 🚀 Launch
+
+> **API**: For more details see the [**related Wiki page**](https://github.com/nicolay-r/bulk-chain/wiki)
+
+```python
+from bulk_chain.core.utils import dynamic_init
+from bulk_chain.api import iter_content
+
+content_it = iter_content(
+    # 1. Your schema.              
+    schema="YOUR_SCHEMA.json",
+    # 2. Your third-party model implementation.
+    llm=dynamic_init(class_filepath="replicate_104.py", class_name="Replicate")(api_token="<API-KEY>"),
+    # 3. Customize your inference and result providing modes: 
+    infer_mode="batch_async", 
+    return_mode="batch",
+    # 4. Your iterator of dictionaries
+    input_dicts_it=YOUR_DATA_IT,
+)
+    
+for content in content_it:
+    # Handle your LLM responses here ...
+```
+
+
+# Embed your LLM
+
+All you have to do is to implement `BaseLM` class, that includes:
+* `__init__` -- for setting up *batching mode support* and (optional) *model name*;
+* `ask(prompt)` -- infer your model with the given `prompt`.
+
+See examples with models [at nlp-thirdgate 🌌](https://github.com/nicolay-r/nlp-thirdgate?tab=readme-ov-file#llm).

bulk_chain-1.1.0/bulk_chain/api.py

@@ -0,0 +1,186 @@
+import asyncio
+import collections
+import logging
+import os
+from itertools import chain
+
+from bulk_chain.core.llm_base import BaseLM
+from bulk_chain.core.service_asyncio import AsyncioService
+from bulk_chain.core.service_batch import BatchIterator
+from bulk_chain.core.service_data import DataService
+from bulk_chain.core.service_dict import DictionaryService
+from bulk_chain.core.service_json import JsonService
+from bulk_chain.core.service_schema import SchemaService
+from bulk_chain.core.utils import attempt_wrapper
+
+
+INFER_MODES = {
+    "single": lambda llm, batch, **kwargs: [llm.ask(prompt) for prompt in batch],
+    "single_stream": lambda llm, batch, **kwargs: [llm.ask_stream(prompt) for prompt in batch],
+    "batch": lambda llm, batch, **kwargs: llm.ask(batch),
+    "batch_async": lambda llm, batch, **kwargs: AsyncioService.run_tasks(
+        batch=batch, async_handler=llm.ask_async, event_loop=kwargs.get("event_loop")
+    ),
+    "batch_stream_async": lambda llm, batch, **kwargs: AsyncioService.run_tasks(
+        batch=batch, async_handler=llm.ask_stream_async, event_loop=kwargs.get("event_loop")
+    ),
+}
+
+
+CWD = os.getcwd()
+
+
+def _iter_batch_prompts(c, batch_content_it, **kwargs):
+    for ind_in_batch, entry in enumerate(batch_content_it):
+        content = DataService.get_prompt_text(
+            prompt=entry[c]["prompt"],
+            data_dict=entry,
+            handle_missed_func=kwargs["handle_missed_value_func"])
+        yield ind_in_batch, content
+
+
+def __handle_agen_to_gen(handle, batch, event_loop):
+    """ This handler provides conversion of the async generator to generator (sync).
+    """
+
+    def __wrap_with_index(async_gens):
+        async def wrapper(index, agen):
+            async for item in agen:
+                yield index, item
+        return [wrapper(i, agen) for i, agen in enumerate(async_gens)]
+
+    agen_list = handle(batch, event_loop=event_loop)
+
+    it = AsyncioService.async_gen_to_iter(
+        gen=AsyncioService.merge_generators(*__wrap_with_index(agen_list)),
+        loop=event_loop)
+
+    for ind_in_batch, chunk in it:
+        yield ind_in_batch, str(chunk)
+
+
+def __handle_gen(handle, batch, event_loop):
+    """ This handler deals with the iteration of each individual element of the batch.
+    """
+
+    def _iter_entry_content(entry):
+        if isinstance(entry, str):
+            yield entry
+        elif isinstance(entry, collections.abc.Iterable):
+            for chunk in map(lambda item: str(item), entry):
+                yield chunk
+        else:
+            raise Exception(f"Non supported type `{type(entry)}` for handling output from batch")
+
+    for ind_in_batch, entry in enumerate(handle(batch, event_loop=event_loop)):
+        for chunk in _iter_entry_content(entry=entry):
+            yield ind_in_batch, chunk
+
+
+def _iter_chunks(p_column, batch_content_it, **kwargs):
+    handler = __handle_agen_to_gen if kwargs["infer_mode"] == "batch_stream_async" else __handle_gen
+    p_batch = [item[p_column] for item in batch_content_it]
+    it = handler(handle=kwargs["handle_batch_func"], batch=p_batch, event_loop=kwargs["event_loop"])
+    for ind_in_batch, chunk in it:
+        yield ind_in_batch, chunk
+
+
+def _infer_batch(batch, batch_ind, schema, return_mode, cols=None, **kwargs):
+    assert (isinstance(batch, list))
+
+    if len(batch) == 0:
+        return batch
+
+    if cols is None:
+        first_item = batch[0]
+        cols = list(first_item.keys()) if cols is None else cols
+
+    for c in cols:
+
+        # Handling prompt column.
+        if c in schema.p2r:
+            content_it = _iter_batch_prompts(c=c, batch_content_it=iter(batch), **kwargs)
+            for ind_in_batch, prompt in content_it:
+                batch[ind_in_batch][c] = prompt
+
+        # Handling column for inference.
+        if c in schema.r2p:
+            content_it = _iter_chunks(p_column=schema.r2p[c], batch_content_it=iter(batch), **kwargs)
+            # Register values.
+            for item in batch:
+                item[c] = []
+            for ind_in_batch, chunk in content_it:
+                # Append batch.
+                batch[ind_in_batch][c].append(chunk)
+                # Returning (optional).
+                if return_mode == "chunk":
+                    global_ind = batch_ind * len(batch) + ind_in_batch
+                    yield [global_ind, c, chunk]
+
+            # Convert content to string.
+            for item in batch:
+                item[c] = "".join(item[c])
+
+    if return_mode == "record":
+        for record in batch:
+            yield record
+
+    if return_mode == "batch":
+        yield batch
+
+
+def iter_content(input_dicts_it, llm, schema, batch_size=1, limit_prompt=None,
+                 infer_mode="batch", return_mode="batch", attempts=1, event_loop=None,
+                 **kwargs):
+    """ This method represent Python API aimed at application of `llm` towards
+        iterator of input_dicts via cache_target that refers to the SQLite using
+        the given `schema`
+    """
+    assert (infer_mode in INFER_MODES.keys())
+    assert (return_mode in ["batch", "chunk", "record"])
+    assert (isinstance(llm, BaseLM))
+
+    # Setup event loop.
+    event_loop = asyncio.get_event_loop_policy().get_event_loop() \
+        if event_loop is None else event_loop
+
+    # Quick initialization of the schema.
+    if isinstance(schema, str):
+        schema = JsonService.read(schema)
+    if isinstance(schema, dict):
+        schema = SchemaService(json_data=schema)
+
+    prompts_it = map(
+        lambda data: DictionaryService.custom_update(src_dict=dict(data), other_dict=schema.cot_args),
+        input_dicts_it
+    )
+
+    handle_batch_func = lambda batch, **handle_kwargs: INFER_MODES[infer_mode](
+        llm,
+        DataService.limit_prompts(batch, limit=limit_prompt),
+        **handle_kwargs
+    )
+
+    # Optional wrapping into attempts.
+    if attempts > 1:
+        # Optional setup of the logger.
+        logger = logging.getLogger(__name__)
+        logging.basicConfig(level=logging.INFO)
+
+        attempt_dec = attempt_wrapper(attempts=attempts,
+                                      delay_sec=kwargs.get("attempt_delay_sec", 1),
+                                      logger=logger)
+        handle_batch_func = attempt_dec(handle_batch_func)
+
+    content_it = (_infer_batch(batch=batch,
+                               batch_ind=batch_ind,
+                               infer_mode=infer_mode,
+                               handle_batch_func=handle_batch_func,
+                               handle_missed_value_func=lambda *_: None,
+                               return_mode=return_mode,
+                               schema=schema,
+                               event_loop=event_loop,
+                               **kwargs)
+                  for batch_ind, batch in enumerate(BatchIterator(prompts_it, batch_size=batch_size)))
+
+    yield from chain.from_iterable(content_it)

bulk_chain-1.1.0/bulk_chain/core/llm_base.py

@@ -0,0 +1,24 @@
+class BaseLM(object):
+
+    def __init__(self, **kwargs):
+        pass
+
+    def ask(self, content):
+        """ Assumes to return str.
+        """
+        raise NotImplemented()
+
+    def ask_stream(self, content):
+        """ Assumes to return generator.
+        """
+        raise NotImplemented()
+
+    async def ask_async(self, prompt):
+        """ Assumes to return co-routine.
+        """
+        raise NotImplemented()
+
+    async def ask_stream_async(self, batch):
+        """ Assumes to return AsyncGenerator.
+        """
+        raise NotImplemented()

bulk_chain-1.1.0/bulk_chain/core/service_asyncio.py

@@ -0,0 +1,65 @@
+import asyncio
+from typing import AsyncGenerator, Any
+
+
+class AsyncioService:
+
+    @staticmethod
+    async def _run_tasks_async(batch, async_handler):
+        tasks = [async_handler(prompt) for prompt in batch]
+        return await asyncio.gather(*tasks)
+
+    @staticmethod
+    async def _run_generator(gen, output_queue, idx):
+        try:
+            async for item in gen:
+                await output_queue.put((idx, item))
+        finally:
+            await output_queue.put((idx, StopAsyncIteration))
+
+
+    @staticmethod
+    def run_tasks(event_loop, **tasks_kwargs):
+        return event_loop.run_until_complete(AsyncioService._run_tasks_async(**tasks_kwargs))
+
+    @staticmethod
+    async def merge_generators(*gens: AsyncGenerator[Any, None]) -> AsyncGenerator[Any, None]:
+
+        output_queue = asyncio.Queue()
+        tasks = [
+            asyncio.create_task(AsyncioService._run_generator(gen, output_queue, idx))
+            for idx, gen in enumerate(gens)
+        ]
+
+        finished = set()
+        while len(finished) < len(tasks):
+            idx, item = await output_queue.get()
+            if item is StopAsyncIteration:
+                finished.add(idx)
+            else:
+                yield item
+
+        for task in tasks:
+            task.cancel()
+
+    @staticmethod
+    def async_gen_to_iter(gen, loop=None):
+        """ This approach is limited. Could be considered as legacy.
+            https://stackoverflow.com/questions/71580727/translating-async-generator-into-sync-one/78573267#78573267
+        """
+
+        loop_created = False
+        if loop is None:
+            loop_created = True
+            loop = asyncio.new_event_loop()
+
+        asyncio.set_event_loop(loop)
+        try:
+            while True:
+                try:
+                    yield loop.run_until_complete(gen.__anext__())
+                except StopAsyncIteration:
+                    break
+        finally:
+            if loop_created:
+                loop.close()

{bulk_chain-0.25.3 → bulk_chain-1.1.0}/bulk_chain/core/service_batch.py

@@ -1,27 +1,8 @@
-class BatchService(object):
-
-    @staticmethod
-    def handle_param_as_batch(batch, src_param, tgt_param, handle_batch_func, handle_entry_func):
-        assert (isinstance(batch, list))
-        assert (isinstance(src_param, str))
-        assert (callable(handle_batch_func))
-
-        _batch = [item[src_param] for item in batch]
-
-        # Do handling for the batch.
-        _handled_batch = handle_batch_func(_batch)
-        assert (isinstance(_handled_batch, list))
-
-        # Apply changes.
-        for i, item in enumerate(batch):
-            item[tgt_param] = handle_entry_func(entry=_handled_batch[i], info={"ind": i, "param": tgt_param})
-
-
 class BatchIterator:
 
     def __init__(self, data_iter, batch_size, end_value=None, filter_func=None):
-        assert(isinstance(batch_size, int) and batch_size > 0)
-        assert(callable(end_value) or end_value is None)
+        assert (isinstance(batch_size, int) and batch_size > 0)
+        assert (callable(end_value) or end_value is None)
         self.__data_iter = data_iter
         self.__index = 0
         self.__batch_size = batch_size

{bulk_chain-0.25.3 → bulk_chain-1.1.0}/bulk_chain/core/utils.py

@@ -1,6 +1,7 @@
 import importlib
 import logging
 import sys
+import time
 from collections import Counter
 from os.path import dirname, join, basename
 
@@ -48,28 +49,6 @@ def iter_params(text):
         beg = pe+1
 
 
-def format_model_name(name):
-    return name.replace("/", "_")
-
-
-def parse_filepath(filepath, default_filepath=None, default_ext=None):
-    """ This is an auxiliary function for handling sources and targets from cmd string.
-    """
-    if filepath is None:
-        return default_filepath, default_ext, None
-    info = filepath.split(":")
-    filepath = info[0]
-    meta = info[1] if len(info) > 1 else None
-    ext = filepath.split('.')[-1] if default_ext is None else default_ext
-    return filepath, ext, meta
-
-
-def handle_table_name(name):
-    return name.\
-        replace('-', '_').\
-        replace('.', "_")
-
-
 def auto_import(name, is_class=False):
     """ Import from the external python packages.
     """
@@ -82,10 +61,10 @@ def auto_import(name, is_class=False):
     return m() if is_class else m
 
 
-def dynamic_init(class_dir, class_filepath, class_name=None):
+def dynamic_init(class_filepath, class_name=None):
 
     # Registering path.
-    target = join(class_dir, dirname(class_filepath))
+    target = join(dirname(class_filepath))
     logger.info(f"Adding sys path for `{target}`")
     sys.path.insert(1, target)
     class_path_list = class_filepath.split('/')
@@ -111,3 +90,21 @@ def optional_limit_iter(it_data, limit=None):
         if limit is not None and counter["returned"] > limit:
             break
         yield data
+
+
+def attempt_wrapper(attempts, delay_sec=1, logger=None):
+    def decorator(func):
+        def wrapper(*args, **kwargs):
+            for i in range(attempts):
+                try:
+                    # Do action.
+                    return func(*args, **kwargs)
+                except Exception as e:
+                    if logger is not None:
+                        logger.info(f"Unable to infer the result. Try {i} out of {attempts}.")
+                        logger.info(e)
+                    if delay_sec is not None:
+                        time.sleep(delay_sec)
+            raise Exception(f"Failed after {attempts} attempts")
+        return wrapper
+    return decorator