bulk-chain 0.24.2__py3-none-any.whl → 0.25.1__py3-none-any.whl

bulk_chain/api.py

@@ -0,0 +1,99 @@
+import os
+from itertools import chain
+
+from bulk_chain.core.llm_base import BaseLM
+from bulk_chain.core.service_batch import BatchIterator, BatchService
+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 dynamic_init, find_by_prefix
+
+INFER_MODES = {
+    "default": lambda llm, prompt, limit_prompt=None: llm.ask_core(
+        prompt[:limit_prompt] if limit_prompt is not None else prompt),
+    "batch": lambda llm, batch, limit_prompt=None: llm.ask_core(
+        DataService.limit_prompts(batch, limit=limit_prompt))
+}
+
+
+CWD = os.getcwd()
+
+
+def _update_batch_content(c, batch, schema, infer_func):
+    assert (isinstance(batch, list))
+    assert (isinstance(c, str))
+
+    if c in schema.p2r:
+        for batch_item in batch:
+            batch_item[c] = DataService.get_prompt_text(prompt=batch_item[c]["prompt"], data_dict=batch_item)
+    if c in schema.r2p:
+        p_column = schema.r2p[c]
+        # This instruction takes a lot of time in a non-batching mode.
+        BatchService.handle_param_as_batch(batch=batch,
+                                           src_param=p_column,
+                                           tgt_param=c,
+                                           handle_func=lambda b: infer_func(b))
+
+
+def _infer_batch(batch, schema, infer_func, cols=None):
+    assert (isinstance(batch, list))
+    assert (callable(infer_func))
+
+    if len(batch) == 0:
+        return batch
+
+    if cols is None:
+        first_item = batch[0]
+        cols = first_item.keys() if cols is None else cols
+
+    for c in cols:
+        _update_batch_content(c=c, batch=batch, schema=schema, infer_func=infer_func)
+
+    return batch
+
+
+def iter_content(input_dicts_it, llm, schema, batch_size=1, return_batch=True, limit_prompt=None):
+    """ 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 (isinstance(llm, BaseLM))
+
+    # 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=data, other_dict=schema.cot_args),
+        input_dicts_it
+    )
+
+    content_it = (_infer_batch(batch=batch,
+                               infer_func=lambda batch: INFER_MODES["batch"](llm, batch, limit_prompt),
+                               schema=schema)
+                  for batch in BatchIterator(prompts_it, batch_size=batch_size))
+
+    yield from content_it if return_batch else chain.from_iterable(content_it)
+
+
+def init_llm(adapter, **model_kwargs):
+    """ This method perform dynamic initialization of LLM from third-party resource.
+    """
+
+    # List of the Supported models and their API wrappers.
+    models_preset = {
+        "dynamic": lambda: dynamic_init(class_dir=CWD, class_filepath=llm_model_name,
+                                        class_name=llm_model_params)(**model_kwargs)
+    }
+
+    # Initialize LLM model.
+    params = adapter.split(':')
+    llm_model_type = params[0]
+    llm_model_name = params[1] if len(params) > 1 else params[-1]
+    llm_model_params = ':'.join(params[2:]) if len(params) > 2 else None
+    llm = find_by_prefix(d=models_preset, key=llm_model_type)()
+
+    return llm, llm_model_name

bulk_chain/core/llm_base.py

@@ -6,29 +6,46 @@ from bulk_chain.core.utils import format_model_name
 
 class BaseLM(object):
 
-    def __init__(self, name, attempts=None, delay_sec=1, enable_log=True, **kwargs):
+    def __init__(self, name=None, attempts=None, delay_sec=1, enable_log=True,
+                 support_batching=False, **kwargs):
+
         self.__name = name
         self.__attempts = 1 if attempts is None else attempts
         self.__delay_sec = delay_sec
+        self.__support_batching = support_batching
 
         if enable_log:
             self.__logger = logging.getLogger(__name__)
             logging.basicConfig(level=logging.INFO)
 
-    def ask_safe(self, prompt):
+    def ask_core(self, batch):
 
         for i in range(self.__attempts):
             try:
-                response = self.ask(prompt)
-                return response
-            except:
+                if self.__support_batching:
+                    # Launch in batch mode.
+                    content = self.ask(batch)
+                else:
+                    # Launch in non-batch mode.
+                    assert len(batch) == 1, "The LM does not support batching," \
+                                            f" while size of the content is {len(batch)} which is not equal 1. " \
+                                            f"Please enable batch-supporting or set required inference settings."
+                    content = batch[0]
+
+                response = self.ask(content)
+
+                # Wrapping into batch the response in the case of non-batching mode.
+                return response if self.__support_batching else [response]
+
+            except Exception as e:
                 if self.__logger is not None:
                     self.__logger.info("Unable to infer the result. Try {} out of {}.".format(i, self.__attempts))
+                    self.__logger.info(e)
                 time.sleep(self.__delay_sec)
 
         raise Exception("Can't infer")
 
-    def ask(self, prompt):
+    def ask(self, content):
         raise NotImplemented()
 
     def name(self):

bulk_chain/core/service_args.py

@@ -12,6 +12,11 @@ class CmdArgsService:
     def iter_arguments(lst):
 
         def __release():
+
+            # We use the True value by default to treat the related parameter as flag.
+            if len(buf) == 0:
+                buf.append(True)
+
             return key, buf if len(buf) > 1 else buf[0]
 
         key = None
@@ -29,18 +34,37 @@ class CmdArgsService:
                 buf.append(a)
 
         # Sharing the remaining params.
-        if len(buf) > 0:
+        if key is not None:
             yield __release()
 
     @staticmethod
-    def partition_list(lst, sep):
+    def __find_suffix_ind(lst, idx_from, end_prefix):
+        for i in range(idx_from, len(lst)):
+            if lst[i].startswith(end_prefix):
+                return i
+        return len(lst)
+
+    @staticmethod
+    def extract_native_args(lst, end_prefix):
+        return lst[:CmdArgsService.__find_suffix_ind(lst, idx_from=0, end_prefix=end_prefix)]
+
+    @staticmethod
+    def find_grouped_args(lst, starts_with, end_prefix):
         """Slices a list in two, cutting on index matching "sep"
         """
-        if sep in lst:
-            idx = lst.index(sep)
-            return (lst[:idx], lst[idx+1:])
-        else:
-            return (lst[:], None)
+
+        # Checking the presence of starts_with.
+        # We have to return empty content in the case of absence starts_with in the lst.
+        if starts_with not in lst:
+            return []
+
+        # Assigning start index.
+        idx_from = lst.index(starts_with) + 1
+
+        # Assigning end index.
+        idx_to = CmdArgsService.__find_suffix_ind(lst, idx_from=idx_from, end_prefix=end_prefix)
+
+        return lst[idx_from:idx_to]
 
     @staticmethod
     def args_to_dict(args):

bulk_chain/core/service_batch.py

@@ -0,0 +1,51 @@
+class BatchService(object):
+
+    @staticmethod
+    def handle_param_as_batch(batch, src_param, tgt_param, handle_func):
+        assert (isinstance(batch, list))
+        assert (isinstance(src_param, str))
+        assert (callable(handle_func))
+
+        _batch = [item[src_param] for item in batch]
+
+        # Do handling for the batch.
+        _handled_batch = handle_func(_batch)
+        assert (isinstance(_handled_batch, list))
+
+        # Apply changes.
+        for i, item in enumerate(batch):
+            item[tgt_param] = _handled_batch[i]
+
+
+class BatchIterator:
+
+    def __init__(self, data_iter, batch_size, end_value=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
+        self.__end_value = end_value
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        buffer = []
+        while True:
+            try:
+                data = next(self.__data_iter)
+            except StopIteration:
+                break
+            buffer.append(data)
+            if len(buffer) == self.__batch_size:
+                break
+
+        if len(buffer) > 0:
+            self.__index += 1
+            return buffer
+
+        if self.__end_value is None:
+            raise StopIteration
+        else:
+            return self.__end_value()

bulk_chain/core/service_data.py

@@ -20,3 +20,7 @@ class DataService(object):
         field_names = list(parse_fields_func(prompt))
         return DataService.compose_prompt_text(
             prompt=prompt, data_dict=data_dict, field_names=field_names)
+
+    @staticmethod
+    def limit_prompts(prompts_list, limit=None):
+        return [p[:limit] if limit is not None else p for p in prompts_list]

bulk_chain/core/service_dict.py

@@ -0,0 +1,10 @@
+class DictionaryService:
+
+    @staticmethod
+    def custom_update(src_dict, other_dict):
+        for k, v in other_dict.items():
+            if k in src_dict:
+                raise Exception(f"The key `{k}` is already defined in both dicts with values: "
+                                f"`{src_dict[k]}` (src) and `{v}` (other)")
+            src_dict[k] = v
+        return src_dict

bulk_chain/core/service_llm.py

@@ -1,8 +1,7 @@
-import logging
-
 from bulk_chain.core.llm_base import BaseLM
 from bulk_chain.core.service_data import DataService
 from bulk_chain.core.utils import iter_params
+from bulk_chain.core.utils_logger import StreamedLogger
 
 
 def pad_str(text, pad):
@@ -23,29 +22,32 @@ def nice_output(text, width, pad=4, remove_new_line=False):
     return text_wrap(content=short_text, width=width, handle_line=lambda line: pad_str(line, pad=pad))
 
 
-def chat_with_lm(lm, chain=None, model_name=None):
+def chat_with_lm(lm, preset_dict=None, chain=None, model_name=None, line_width=80, pad=0):
     assert (isinstance(lm, BaseLM))
     assert (isinstance(chain, list))
     assert (isinstance(model_name, str) or model_name is None)
 
-    logger = logging.getLogger(__name__)
-    logging.basicConfig(level=logging.INFO)
+    preset_dict = {} if preset_dict is None else preset_dict
+
+    streamed_logger = StreamedLogger(__name__)
 
     do_exit = False
     model_name = model_name if model_name is not None else "agent"
 
     while not do_exit:
 
-        logger.info("----------------")
+        streamed_logger.info("----------------")
+        streamed_logger.info("\n")
 
         # Launching the CoT engine loop.
-        data_dict = {}
-        for prompt_args in chain:
+        data_dict = {} | preset_dict
+        for chain_ind, prompt_args in enumerate(chain):
 
             # Processing the prompt.
             prompt = prompt_args["prompt"]
 
             # Filling necessary parameters.
+            user_informed = False
             field_names = list(iter_params(prompt))
             for ind, f_name in enumerate(field_names):
 
@@ -54,6 +56,7 @@ def chat_with_lm(lm, chain=None, model_name=None):
 
                 user_input = input(f"Enter your prompt for `{f_name}` ({ind+1}/{len(field_names)}) "
                                    f"(or 'exit' to quit): ")
+                user_informed = True
 
                 if user_input.lower() == 'exit':
                     do_exit = True
@@ -64,19 +67,37 @@ def chat_with_lm(lm, chain=None, model_name=None):
             if do_exit:
                 break
 
+            # In the case of the initial interaction with the chain.
+            # we make sure that aware user for starting interaction.
+            if chain_ind == 0 and not user_informed:
+                user_input = input(f"Enter to continue (or 'exit' to quit) ...")
+                if user_input.lower() == 'exit':
+                    do_exit = True
+
             # Finally asking LLM.
             DataService.compose_prompt_text(prompt=prompt, data_dict=data_dict, field_names=field_names)
             actual_prompt = DataService.get_prompt_text(prompt=prompt, data_dict=data_dict)
 
             # Returning meta information, passed to LLM.
-            pad = 4
-            logger.info(pad_str(f"{model_name} (ask) ->", pad=pad))
-            logger.info(nice_output(actual_prompt, pad=pad*2, remove_new_line=True, width=80))
+            streamed_logger.info(pad_str(f"{model_name} (ask [{chain_ind+1}/{len(chain)}]) ->", pad=pad))
+            streamed_logger.info("\n")
+            streamed_logger.info(nice_output(actual_prompt, pad=pad, remove_new_line=True, width=line_width))
+            streamed_logger.info("\n\n")
 
             # Response.
-            response = lm.ask_safe(actual_prompt)
-            logger.info(pad_str(f"{model_name} (resp)->", pad=pad))
-            logger.info(nice_output(response, pad=pad*2, remove_new_line=False, width=80))
+            response = lm.ask_core(batch=[actual_prompt])[0]
+            streamed_logger.info(pad_str(f"{model_name} (resp [{chain_ind+1}/{len(chain)}])->", pad=pad))
+            streamed_logger.info("\n")
+            if isinstance(response, str):
+                streamed_logger.info(nice_output(response, pad=pad, remove_new_line=False, width=line_width))
+                buffer = [response]
+            else:
+                buffer = []
+                for chunk in response:
+                    streamed_logger.info(chunk)
+                    buffer.append(str(chunk))
+
+            streamed_logger.info("\n\n")
 
             # Collecting the answer for the next turn.
-            data_dict[prompt_args["out"]] = response
+            data_dict[prompt_args["out"]] = "".join(buffer)

bulk_chain/core/service_schema.py

@@ -2,12 +2,11 @@ class SchemaService(object):
 
     def __init__(self, json_data):
         self.src = json_data
-        self.name = self.src["name"]
         self.r2p, self.p2r, self.cot_args, self.chain = SchemaService.__init_schema(prompts=json_data["schema"])
 
     @classmethod
     def from_prompt(cls, prompt):
-        prompt_schema = {"name": "prompt", "schema": [{"prompt": prompt, "out": "response", "in": "prompt"}]}
+        prompt_schema = {"schema": [{"prompt": prompt, "out": "response", "in": "prompt"}]}
         return cls(prompt_schema)
 
     @staticmethod

bulk_chain/core/utils_logger.py

@@ -0,0 +1,41 @@
+import logging
+
+
+def StreamedLogger(name: str) -> logging.Logger:
+    """ https://medium.com/@r.das699/optimizing-logging-practices-for-streaming-data-in-python-521798e1ed82
+    """
+    root_handlers = logging.getLogger().handlers
+    current_logger = logging.getLogger(name)
+    if not root_handlers:
+        new_handler = logging.StreamHandler()
+        new_handler.terminator = ""
+        new_handler.setFormatter(logging.Formatter("%(message)s"))
+        current_logger.addHandler(new_handler)
+        current_logger.propagate = False
+        current_logger.setLevel(logging.INFO)
+        return current_logger
+
+    for handler in current_logger.handlers[:]:
+        current_logger.removeHandler(handler)
+
+    for handler_r in root_handlers:
+        if type(handler_r) is logging.StreamHandler:
+            new_handler = logging.StreamHandler()
+            new_handler.terminator = ""
+            new_handler.setFormatter(logging.Formatter("%(message)s"))
+            current_logger.addHandler(new_handler)
+        elif type(handler_r) is logging.FileHandler:
+            new_handler = logging.FileHandler(
+                handler_r.baseFilename,
+                handler_r.mode,
+                handler_r.encoding,
+                handler_r.delay,
+                handler_r.errors,
+            )
+            new_handler.terminator = ""  # This will stop the printing in new line
+            new_handler.setFormatter(logging.Formatter("%(message)s"))
+            current_logger.addHandler(new_handler)
+        else:
+            continue
+    current_logger.propagate = False  # Don't propagate to root logger
+    return current_logger

bulk_chain/demo.py

@@ -0,0 +1,85 @@
+import json
+
+import argparse
+import logging
+import sys
+
+from source_iter.service_jsonl import JsonlService
+
+from bulk_chain.api import init_llm
+from bulk_chain.core.service_args import CmdArgsService
+from bulk_chain.core.service_json import JsonService
+from bulk_chain.core.service_llm import chat_with_lm
+from bulk_chain.core.service_schema import SchemaService
+from bulk_chain.core.utils import parse_filepath
+
+logger = logging.getLogger(__name__)
+logging.basicConfig(level=logging.INFO)
+
+
+def iter_from_json(filepath):
+    with open(filepath, "r") as f:
+        content = json.load(f)
+        for key, value in content.items():
+            yield key, value
+
+
+def iter_from_text_file(filepath):
+    with open(filepath, "r") as f:
+        yield filepath.split('.')[0], f.read()
+
+
+if __name__ == '__main__':
+
+    parser = argparse.ArgumentParser(description="LLM demo usage based on CoT schema")
+    parser.add_argument('--adapter', dest='adapter', type=str, default=None)
+    parser.add_argument('--attempts', dest='attempts', type=int, default=None)
+    parser.add_argument('--src', dest='src', type=str, nargs="*", default=None)
+    parser.add_argument('--schema', dest='schema', type=str, default=None,
+                        help="Path to the JSON file that describes schema")
+    parser.add_argument('--limit-prompt', dest="limit_prompt", type=int, default=None,
+                        help="Optional trimming prompt by the specified amount of characters.")
+
+    # Extract native arguments.
+    native_args = CmdArgsService.extract_native_args(sys.argv, end_prefix="%%")
+    args = parser.parse_args(args=native_args[1:])
+
+    # Extract model-related arguments and Initialize Large Language Model.
+    model_args = CmdArgsService.find_grouped_args(lst=sys.argv, starts_with="%%m", end_prefix="%%")
+    model_args_dict = CmdArgsService.args_to_dict(model_args) | {"attempts": args.attempts}
+    llm, llm_model_name = init_llm(adapter=args.adapter, **model_args_dict)
+
+    # Setup schema.
+    schema = SchemaService(json_data=JsonService.read(args.schema))
+    schema_name = schema.src.get("name", None)
+    if schema is not None:
+        logger.info(f"Using schema: {schema_name}")
+
+    output_providers = {
+        "jsonl": lambda filepath, data_it, header:
+            JsonlService.write(target=filepath,
+                               data_it=map(lambda item: {key: item[i] for i, key in enumerate(header)}, data_it))
+    }
+
+    input_file_handlers = {
+        "json": lambda filepath: iter_from_json(filepath),
+        "txt": lambda filepath: iter_from_text_file(filepath)
+    }
+
+    # Input extension type defines the provider.
+    if args.src is None:
+        args.src = []
+    if isinstance(args.src, str):
+        args.src = [args.src]
+    sources = [parse_filepath(s) for s in args.src]
+
+    preset_dict = {}
+    for fp, ext, _ in sources:
+        for key, value in input_file_handlers[ext](fp):
+            if key in preset_dict:
+                raise Exception(f"While at handling {fp}: Key {key} is already registered!")
+            preset_dict[key] = value
+
+    # Launch Demo.
+    chat_with_lm(llm, preset_dict=preset_dict, chain=schema.chain, model_name=llm_model_name,
+                 line_width=120)

bulk_chain/infer.py

@@ -1,98 +1,67 @@
-import os
 from os.path import join, basename
 
 import argparse
 import logging
 import sys
 
-from tqdm import tqdm
-
 from source_iter.service_csv import CsvService
 from source_iter.service_jsonl import JsonlService
 from source_iter.service_sqlite import SQLite3Service
+from tqdm import tqdm
 
+from bulk_chain.api import INFER_MODES, _infer_batch, CWD, init_llm
 from bulk_chain.core.llm_base import BaseLM
 from bulk_chain.core.service_args import CmdArgsService
-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_llm import chat_with_lm
 from bulk_chain.core.service_schema import SchemaService
-from bulk_chain.core.utils import dynamic_init, find_by_prefix, handle_table_name, optional_limit_iter, parse_filepath
+from bulk_chain.core.utils import handle_table_name, optional_limit_iter, parse_filepath
 
 logger = logging.getLogger(__name__)
 logging.basicConfig(level=logging.INFO)
 
+WRITER_PROVIDERS = {
+    "sqlite": lambda filepath, table_name, data_it, infer_data_func, **kwargs: SQLite3Service.write(
+        data_it=data_it, target=filepath, table_name=table_name, data2col_func=infer_data_func,
+        skip_existed=True, **kwargs)
+}
 
-CWD = os.getcwd()
-
-
-def init_llm(**model_kwargs):
-    """ This method perform dynamic initialization of LLM from third-party resource.
-    """
-
-    # List of the Supported models and their API wrappers.
-    models_preset = {
-        "dynamic": lambda: dynamic_init(class_dir=CWD, class_filepath=llm_model_name,
-                                        class_name=llm_model_params)(**model_kwargs)
-    }
-
-    # Initialize LLM model.
-    params = args.adapter.split(':')
-    llm_model_type = params[0]
-    llm_model_name = params[1] if len(params) > 1 else params[-1]
-    llm_model_params = ':'.join(params[2:]) if len(params) > 2 else None
-    llm = find_by_prefix(d=models_preset, key=llm_model_type)()
-
-    return llm, llm_model_name
-
+READER_PROVIDERS = {
+    "sqlite": lambda filepath, table_name: SQLite3Service.read(filepath, table=table_name)
+}
 
-def init_schema(json_filepath):
-    return SchemaService(json_data=JsonService.read(json_filepath))
 
-
-def iter_content(input_dicts_iter, llm, schema, cache_target, cache_table, id_column_name):
-    """ 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`
-    """
+def iter_content_cached(input_dicts_it, llm, schema, cache_target, limit_prompt=None, **cache_kwargs):
     assert (isinstance(llm, BaseLM))
-    assert (isinstance(schema, SchemaService))
     assert (isinstance(cache_target, str))
-    assert (isinstance(cache_table, str))
-
-    infer_modes = {
-        "default": lambda prompt: llm.ask_safe(prompt[:args.limit_prompt] if args.limit_prompt is not None else prompt)
-    }
-
-    def optional_update_data_records(c, data):
-        assert (isinstance(c, str))
-
-        if c in schema.p2r:
-            data[c] = DataService.get_prompt_text(prompt=data[c]["prompt"], data_dict=data)
-        if c in schema.r2p:
-            p_column = schema.r2p[c]
-            # This instruction takes a lot of time in a non-batching mode.
-            data[c] = infer_modes["default"](data[p_column])
 
-        return data[c]
+    # Quick initialization of the schema.
+    if isinstance(schema, str):
+        schema = JsonService.read(schema)
+    if isinstance(schema, dict):
+        schema = SchemaService(json_data=schema)
 
-    cache_providers = {
-        "sqlite": lambda filepath, table_name, data_it: SQLite3Service.write_missed(
-            data_it=data_it, target=filepath,
-            data2col_func=optional_update_data_records,
-            table_name=handle_table_name(table_name if table_name is not None else "contents"),
-            id_column_name=id_column_name)
-    }
+    # Iterator of the queries.
+    prompts_it = map(
+        lambda data: DictionaryService.custom_update(src_dict=data, other_dict=schema.cot_args),
+        input_dicts_it
+    )
 
-    # We optionally wrap into limiter.
-    queries_it = optional_limit_iter(
-        it_data=map(lambda data: data.update(schema.cot_args) or data, input_dicts_iter),
-        limit=args.limit)
+    # Parse target.
+    cache_filepath, _, cache_table = parse_filepath(filepath=cache_target)
 
-    # Provide data caching.
-    cache_providers["sqlite"](cache_target, table_name=tgt_meta, data_it=tqdm(queries_it, desc="Iter content"))
+    # Perform caching first.
+    WRITER_PROVIDERS["sqlite"](
+        filepath=cache_filepath, table_name=cache_table,
+        data_it=tqdm(prompts_it, desc="Iter content"),
+        infer_data_func=lambda c, prompt: _infer_batch(
+            batch=[prompt], cols=[c],
+            infer_func=lambda batch: INFER_MODES["default"](llm, batch, limit_prompt),
+            schema=schema)[0][c],
+        **cache_kwargs)
 
-    return SQLite3Service.read(cache_target, table=cache_table)
+    # Then retrieve data.
+    return READER_PROVIDERS["sqlite"](filepath=cache_filepath, table_name=cache_table)
 
 
 if __name__ == '__main__':
@@ -101,7 +70,7 @@ if __name__ == '__main__':
     parser.add_argument('--adapter', dest='adapter', type=str, default=None)
     parser.add_argument('--attempts', dest='attempts', type=int, default=None)
     parser.add_argument('--id-col', dest='id_col', type=str, default="uid")
-    parser.add_argument('--src', dest='src', type=str, default=None)
+    parser.add_argument('--src', dest='src', type=str, nargs="?", default=None)
     parser.add_argument('--schema', dest='schema', type=str, default=None,
                         help="Path to the JSON file that describes schema")
     parser.add_argument('--to', dest='to', type=str, default=None, choices=["csv", "sqlite"])
@@ -111,31 +80,48 @@ if __name__ == '__main__':
     parser.add_argument('--limit-prompt', dest="limit_prompt", type=int, default=None,
                         help="Optional trimming prompt by the specified amount of characters.")
 
-    native_args, model_args = CmdArgsService.partition_list(lst=sys.argv, sep="%%")
-
+    # Extract native arguments.
+    native_args = CmdArgsService.extract_native_args(sys.argv, end_prefix="%%")
     args = parser.parse_args(args=native_args[1:])
 
-    # Initialize Large Language Model.
+    # Extract csv-related arguments.
+    csv_args = CmdArgsService.find_grouped_args(lst=sys.argv, starts_with="%%csv", end_prefix="%%")
+    csv_args_dict = CmdArgsService.args_to_dict(csv_args)
+
+    # Extract model-related arguments and Initialize Large Language Model.
+    model_args = CmdArgsService.find_grouped_args(lst=sys.argv, starts_with="%%m", end_prefix="%%")
     model_args_dict = CmdArgsService.args_to_dict(model_args) | {"attempts": args.attempts}
-    llm, llm_model_name = init_llm(**model_args_dict)
+    llm, llm_model_name = init_llm(adapter=args.adapter, **model_args_dict)
 
     # Setup schema.
-    schema = init_schema(args.schema)
+    schema = SchemaService(json_data=JsonService.read(args.schema))
+    schema_name = schema.src.get("name", None)
     if schema is not None:
-        logger.info(f"Using schema: {schema.name}")
+        logger.info(f"Using schema: {schema_name}")
 
     input_providers = {
-        None: lambda _: chat_with_lm(llm, chain=schema.chain, model_name=llm_model_name),
         "csv": lambda filepath: CsvService.read(src=filepath, row_id_key=args.id_col,
                                                 as_dict=True, skip_header=True,
-                                                delimiter=model_args_dict.get("delimiter", "\t"),
-                                                escapechar=model_args_dict.get("escapechar", None)),
+                                                delimiter=csv_args_dict.get("delimiter", ","),
+                                                escapechar=csv_args_dict.get("escapechar", None)),
+        "tsv": lambda filepath: CsvService.read(src=filepath, row_id_key=args.id_col,
+                                                as_dict=True, skip_header=True,
+                                                delimiter=csv_args_dict.get("delimiter", "\t"),
+                                                escapechar=csv_args_dict.get("escapechar", None)),
         "jsonl": lambda filepath: JsonlService.read(src=filepath, row_id_key=args.id_col)
     }
 
     output_providers = {
-        "csv": lambda filepath, data_it, header:
-            CsvService.write(target=filepath, data_it=data_it, header=header, it_type=None),
+        "csv": lambda filepath, data_it, header: CsvService.write(target=filepath,
+                                                                  data_it=data_it, header=header,
+                                                                  delimiter=csv_args_dict.get("delimiter", ","),
+                                                                  escapechar=csv_args_dict.get("escapechar", None),
+                                                                  it_type=None),
+        "tsv": lambda filepath, data_it, header: CsvService.write(target=filepath,
+                                                                  data_it=data_it, header=header,
+                                                                  delimiter=csv_args_dict.get("delimiter", "\t"),
+                                                                  escapechar=csv_args_dict.get("escapechar", None),
+                                                                  it_type=None),
         "jsonl": lambda filepath, data_it, header:
         JsonlService.write(target=filepath,
                            data_it=map(lambda item: {key: item[i] for i, key in enumerate(header)}, data_it))
@@ -145,32 +131,32 @@ if __name__ == '__main__':
     args.output = args.output.format(model=llm.name()) if args.output is not None else args.output
     tgt_filepath, tgt_ext, tgt_meta = parse_filepath(args.output, default_ext=args.to)
 
-    # Input extension type defines the provider.
+    # We do not support multiple files for other modes.
     src_filepath, src_ext, src_meta = parse_filepath(args.src)
 
-    # Check whether we are in chat mode.
-    if src_ext is None:
-        input_providers[src_ext](None)
-        exit(0)
+    def default_output_file_template(ext):
+        # This is a default template for output files to be generated.
+        return "".join(["_".join([join(CWD, basename(src_filepath)), llm.name(), schema_name]), ext])
 
     # Setup cache target as well as the related table.
-    cache_target = "".join(["_".join([join(CWD, basename(src_filepath)), llm.name(), schema.name]), f".sqlite"]) \
-        if tgt_filepath is None else tgt_filepath
+    cache_filepath = default_output_file_template(".sqlite") if tgt_filepath is None else tgt_filepath
     cache_table = handle_table_name(tgt_meta if tgt_meta is not None else "contents")
 
-    data_it = iter_content(input_dicts_iter=input_providers[src_ext](src_filepath),
-                           schema=schema,
-                           llm=llm,
-                           id_column_name=args.id_col,
-                           cache_target=cache_target,
-                           cache_table=cache_table)
+    # This is a content that we extracted via input provider.
+    it_data = input_providers[src_ext](src_filepath)
+
+    data_it = iter_content_cached(input_dicts_it=optional_limit_iter(it_data=it_data, limit=args.limit),
+                                  limit_prompt=args.limit_prompt,
+                                  schema=schema,
+                                  llm=llm,
+                                  id_column_name=args.id_col,
+                                  cache_target=":".join([cache_filepath, cache_table]))
 
     # Setup output target
     tgt_ext = src_ext if tgt_ext is None else tgt_ext
-    output_target = "".join(["_".join([join(CWD, basename(src_filepath)), llm.name(), schema.name]), f".{tgt_ext}"]) \
-        if tgt_filepath is None else tgt_filepath
+    output_target = default_output_file_template(f".{tgt_ext}") if tgt_filepath is None else tgt_filepath
 
     # Perform output writing process.
     output_providers[tgt_ext](filepath=output_target,
                               data_it=data_it,
-                              header=SQLite3Service.read_columns(target=cache_target, table=cache_table))
+                              header=SQLite3Service.read_columns(target=cache_filepath, table=cache_table))

bulk_chain-0.25.1.dist-info/METADATA

@@ -0,0 +1,131 @@
+Metadata-Version: 2.1
+Name: bulk_chain
+Version: 0.25.1
+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
+Author-email: rusnicolay@gmail.com
+License: MIT License
+Keywords: natural language processing,chain-of-thought,reasoning
+Classifier: Programming Language :: Python
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Topic :: Software Development :: Libraries :: Python Modules
+Classifier: Topic :: Scientific/Engineering :: Information Analysis
+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 0.25.1
+![](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>
+</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** served in `CSV`/`JSONL`.
+
+### Extra Features
+* ✅ **Progress caching [for remote LLMs]**: withstanding exception during LLM calls by using `sqlite3` engine for caching LLM answers;
+
+
+# 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
+
+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).
+
+## API
+
+Please take a look at the [**related Wiki page**](https://github.com/nicolay-r/bulk-chain/wiki)
+
+## Shell
+
+> **NOTE:** You have to install `source-iter` and `tqdm` packages that actual [dependencies](dependencies.txt) of this project
+
+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>"
+```
+
+Or, you can launch **demo mode** to interact with LLM via command line: 
+> **NOTE:** Demo supports streaming!
+```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
+```
+
+# 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-0.25.1.dist-info/RECORD

@@ -0,0 +1,20 @@
+bulk_chain/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+bulk_chain/api.py,sha256=3q1t4A5wop_BRgYanFCCSQBiGu38P9ds0hTbuxNIUKQ,3590
+bulk_chain/demo.py,sha256=x5OPyc7NSSr9KsfSx09KQkntJIev6gytqdhtx4b9FdU,3224
+bulk_chain/infer.py,sha256=VsqP8CP1pHBNeaDYAv__W6EPL8qBibuAtHzl4-S54iU,8148
+bulk_chain/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+bulk_chain/core/llm_base.py,sha256=uX_uibm5y8STfMKNYL64EeF8UowfJGwCD_t-uftHoJE,1849
+bulk_chain/core/service_args.py,sha256=lq4Veuh4QNu8mlCv8MT9S1rMxTn4FKalyp-3boYonVk,2136
+bulk_chain/core/service_batch.py,sha256=yQr6fbQd4ifQBGMhZMrQQeZpXtDchMKMGJi8XPG7thc,1430
+bulk_chain/core/service_data.py,sha256=ZjJDtd1jrQm9hRCXMqe4CT_qF2XDbWBE1lVibP7tAWo,942
+bulk_chain/core/service_dict.py,sha256=lAghLU-3V3xYGv5BTA327Qcw8UJYmgQRMFdggzlrUgo,383
+bulk_chain/core/service_json.py,sha256=6o1xM_8c9QEjH9Q3qEmJylU9nahfRXhUd5sFF2dGJwo,182
+bulk_chain/core/service_llm.py,sha256=uxtEw9OANFyjS4IECos_N5pUfAPQZVul61m64gnlwt0,3880
+bulk_chain/core/service_schema.py,sha256=KIP4n0Tz2h1i7SIMGhgAhoiCgUFXOT1rzMt38yACS2U,1154
+bulk_chain/core/utils.py,sha256=UV6Cefaw7yZiYblsCr-s9LsbcI83xe7eESBvha9A2Og,2784
+bulk_chain/core/utils_logger.py,sha256=BD-ADxaeeuHztaYjqtIY_cIzc5r2Svq9XwRtrgIEqyI,1636
+bulk_chain-0.25.1.dist-info/LICENSE,sha256=VF9SjNpwwSSFEY_eP_8A1ocDCrbwfjI1pZexXdCkOwo,1076
+bulk_chain-0.25.1.dist-info/METADATA,sha256=EQNb6FwovXUSNXL5QpWf-3003863kL84ZWU7wya3ZeM,5071
+bulk_chain-0.25.1.dist-info/WHEEL,sha256=pL8R0wFFS65tNSRnaOVrsw9EOkOqxLrlUPenUYnJKNo,91
+bulk_chain-0.25.1.dist-info/top_level.txt,sha256=Hxq_wyH-GDXKBaA63UfBIiMJO2eCHJG5EOrXDphpeB4,11
+bulk_chain-0.25.1.dist-info/RECORD,,

bulk_chain-0.24.2.dist-info/METADATA

@@ -1,98 +0,0 @@
-Metadata-Version: 2.1
-Name: bulk_chain
-Version: 0.24.2
-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
-Author-email: rusnicolay@gmail.com
-License: MIT License
-Keywords: natural language processing,chain-of-thought,reasoning
-Classifier: Programming Language :: Python
-Classifier: Programming Language :: Python :: 3.9
-Classifier: Topic :: Software Development :: Libraries :: Python Modules
-Classifier: Topic :: Scientific/Engineering :: Information Analysis
-Classifier: Topic :: Text Processing :: Linguistic
-Requires-Python: >=3.6
-Description-Content-Type: text/markdown
-License-File: LICENSE
-Requires-Dist: tqdm
-Requires-Dist: source-iter ==0.24.2
-
-# bulk-chain 0.24.2
-![](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)
-
-<p align="center">
-    <img src="logo.png"/>
-</p>
-
-A lightweight, no-strings-attached **[Chain-of-Thought](https://arxiv.org/abs/2201.11903) framework** for your LLM, ensuring reliable results for bulk input requests stored in `CSV` / `JSONL` / `sqlite`.
-It allows applying series of prompts formed into `schema` (See [related section](#chain-of-thought-schema))
-
-### Features
-* ✅ **No-strings**: you're free to LLM dependencies and flexible `venv` customization.
-* ✅ **Provides iterator over infinite amount of input contexts** served in `CSV`/`JSONL`.
-* ✅ **Progress caching**: withstanding exception during LLM calls by using `sqlite3` engine for caching LLM answers;
-* ✅ **Support schemas descriptions** for Chain-of-Thought concept.
-
-# Installation
-
-```bash
-pip install bulk-chain
-```
-
-## 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"},
-]
-}
-```
-
-Another templates are available [here](/ext/schema/thor_cot_schema.json).
-
-# Usage
-
-Just **three** simple steps:
-
-1. Define your [CoT Schema](#chain-of-thought-schema), or fetch it as shown below:
-```bash
-!wget https://raw.githubusercontent.com/nicolay-r/bulk-chain/refs/heads/master/ext/schema/default.json
-```
-2. Fetch or write your own **model** or pick the one [preset here](/ext/):
-```bash
-!wget https://raw.githubusercontent.com/nicolay-r/bulk-chain/refs/heads/master/ext/flan_t5.py
-```
-
-3. Launch inference in (chat mode):
-```bash
-!python -m bulk_chain.infer \
-    --schema "default.json" \
-    --adapter "dynamic:flan_t5.py:FlanT5" \
-    %% \
-    --device "cpu" \
-    --temp 0.1
-```
-
-# Embed your LLM
-
-All you have to do is to implement `BaseLM` class, that includes:
-* `__init__` -- for initialization;
-* `ask(prompt)` -- infer your model with the given `prompt`.
-
-See examples with models [here](/ext).

bulk_chain-0.24.2.dist-info/RECORD

@@ -1,15 +0,0 @@
-bulk_chain/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-bulk_chain/infer.py,sha256=QgbR64A1JS8B9oh0_ruynEfdCpoG1rPHVMtk5Z0Ch2U,7476
-bulk_chain/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-bulk_chain/core/llm_base.py,sha256=Cakuuy4jTVPOta5TyEPFTPGvFpJfM6m0dAIAYSu7zFM,1008
-bulk_chain/core/service_args.py,sha256=Qr3rHsAB8wnajB-DbU-GjiEpRZFP4D6s1lVTpLkPPX4,1294
-bulk_chain/core/service_data.py,sha256=18gQwSCTEsI7XFukq8AE5lDJX_QQRpasaH69g6EddV0,797
-bulk_chain/core/service_json.py,sha256=6o1xM_8c9QEjH9Q3qEmJylU9nahfRXhUd5sFF2dGJwo,182
-bulk_chain/core/service_llm.py,sha256=NoD5KHGtXCmN8SlpgH0Z5KCmSxZcnVVfp65vhVRoG84,2742
-bulk_chain/core/service_schema.py,sha256=JVhOv2YP5VEtiwOq_zgCzhS2uF_BOATAgg6fmKRf2NQ,1209
-bulk_chain/core/utils.py,sha256=UV6Cefaw7yZiYblsCr-s9LsbcI83xe7eESBvha9A2Og,2784
-bulk_chain-0.24.2.dist-info/LICENSE,sha256=VF9SjNpwwSSFEY_eP_8A1ocDCrbwfjI1pZexXdCkOwo,1076
-bulk_chain-0.24.2.dist-info/METADATA,sha256=yEKF0X90AvNw6yq-W7oBdTLiH8KSrCKRHaERhOFLXFA,3685
-bulk_chain-0.24.2.dist-info/WHEEL,sha256=pL8R0wFFS65tNSRnaOVrsw9EOkOqxLrlUPenUYnJKNo,91
-bulk_chain-0.24.2.dist-info/top_level.txt,sha256=Hxq_wyH-GDXKBaA63UfBIiMJO2eCHJG5EOrXDphpeB4,11
-bulk_chain-0.24.2.dist-info/RECORD,,