trustgraph/trustgraph-flow/trustgraph/chunking/token/chunker.py
2025-10-06 17:54:26 +01:00

128 lines
3.4 KiB
Python
Executable file

"""
Simple decoder, accepts text documents on input, outputs chunks from the
as text as separate output objects.
"""
import logging
from langchain_text_splitters import TokenTextSplitter
from prometheus_client import Histogram
from ... schema import TextDocument, Chunk
from ... base import ChunkingService, ConsumerSpec, ProducerSpec
# Module logger
logger = logging.getLogger(__name__)
default_ident = "chunker"
class Processor(ChunkingService):
def __init__(self, **params):
id = params.get("id", default_ident)
chunk_size = params.get("chunk_size", 250)
chunk_overlap = params.get("chunk_overlap", 15)
super(Processor, self).__init__(
**params | { "id": id }
)
# Store default values for parameter override
self.default_chunk_size = chunk_size
self.default_chunk_overlap = chunk_overlap
if not hasattr(__class__, "chunk_metric"):
__class__.chunk_metric = Histogram(
'chunk_size', 'Chunk size',
["id", "flow"],
buckets=[100, 160, 250, 400, 650, 1000, 1600,
2500, 4000, 6400, 10000, 16000]
)
self.text_splitter = TokenTextSplitter(
encoding_name="cl100k_base",
chunk_size=chunk_size,
chunk_overlap=chunk_overlap,
)
self.register_specification(
ConsumerSpec(
name = "input",
schema = TextDocument,
handler = self.on_message,
)
)
self.register_specification(
ProducerSpec(
name = "output",
schema = Chunk,
)
)
logger.info("Token chunker initialized")
async def on_message(self, msg, consumer, flow):
v = msg.value()
logger.info(f"Chunking document {v.metadata.id}...")
# Extract chunk parameters from flow (allows runtime override)
chunk_size, chunk_overlap = await self.chunk_document(
msg, consumer, flow,
self.default_chunk_size,
self.default_chunk_overlap
)
# Create text splitter with effective parameters
text_splitter = TokenTextSplitter(
encoding_name="cl100k_base",
chunk_size=chunk_size,
chunk_overlap=chunk_overlap,
)
texts = text_splitter.create_documents(
[v.text.decode("utf-8")]
)
for ix, chunk in enumerate(texts):
logger.debug(f"Created chunk of size {len(chunk.page_content)}")
r = Chunk(
metadata=v.metadata,
chunk=chunk.page_content.encode("utf-8"),
)
__class__.chunk_metric.labels(
id=consumer.id, flow=consumer.flow
).observe(len(chunk.page_content))
await flow("output").send(r)
logger.debug("Document chunking complete")
@staticmethod
def add_args(parser):
ChunkingService.add_args(parser)
parser.add_argument(
'-z', '--chunk-size',
type=int,
default=250,
help=f'Chunk size (default: 250)'
)
parser.add_argument(
'-v', '--chunk-overlap',
type=int,
default=15,
help=f'Chunk overlap (default: 15)'
)
def run():
Processor.launch(default_ident, __doc__)