feat(ingest): Created a faster ingestion mode - pipeline (#1750)

* Unify pgvector and postgres connection settings

* Remove local changes

* Update file pgvector->postgres

* postgresql should be postgres

* Adding pipeline ingestion mode

* disable hugging face parallelism.  Continue on file to doc transform failure

* Semaphore to limit docq async workers. ETA reporting

fern/docs/pages/manual/ingestion.mdx

@@ -62,6 +62,7 @@ The following ingestion mode exist:
 * `simple`: historic behavior, ingest one document at a time, sequentially
 * `batch`: read, parse, and embed multiple documents using batches (batch read, and then batch parse, and then batch embed)
 * `parallel`: read, parse, and embed multiple documents in parallel. This is the fastest ingestion mode for local setup.
+* `pipeline`: Alternative to parallel.
 To change the ingestion mode, you can use the `embedding.ingest_mode` configuration value. The default value is `simple`.
 
 To configure the number of workers used for parallel or batched ingestion, you can use

private_gpt/components/ingest/ingest_component.py

@@ -6,6 +6,7 @@ import multiprocessing.pool
 import os
 import threading
 from pathlib import Path
+from queue import Queue
 from typing import Any
 
 from llama_index.core.data_structs import IndexDict
@@ -13,12 +14,13 @@ from llama_index.core.embeddings.utils import EmbedType
 from llama_index.core.indices import VectorStoreIndex, load_index_from_storage
 from llama_index.core.indices.base import BaseIndex
 from llama_index.core.ingestion import run_transformations
-from llama_index.core.schema import Document, TransformComponent
+from llama_index.core.schema import BaseNode, Document, TransformComponent
 from llama_index.core.storage import StorageContext
 
 from private_gpt.components.ingest.ingest_helper import IngestionHelper
 from private_gpt.paths import local_data_path
 from private_gpt.settings.settings import Settings
+from private_gpt.utils.eta import eta
 
 logger = logging.getLogger(__name__)
 
@@ -314,6 +316,170 @@ class ParallelizedIngestComponent(BaseIngestComponentWithIndex):
         self._file_to_documents_work_pool.terminate()
 
 
+class PipelineIngestComponent(BaseIngestComponentWithIndex):
+    """Pipeline ingestion - keeping the embedding worker pool as busy as possible.
+
+    This class implements a threaded ingestion pipeline, which comprises two threads
+    and two queues. The primary thread is responsible for reading and parsing files
+    into documents. These documents are then placed into a queue, which is
+    distributed to a pool of worker processes for embedding computation. After
+    embedding, the documents are transferred to another queue where they are
+    accumulated until a threshold is reached. Upon reaching this threshold, the
+    accumulated documents are flushed to the document store, index, and vector
+    store.
+
+    Exception handling ensures robustness against erroneous files. However, in the
+    pipelined design, one error can lead to the discarding of multiple files. Any
+    discarded files will be reported.
+    """
+
+    NODE_FLUSH_COUNT = 5000  # Save the index every # nodes.
+
+    def __init__(
+        self,
+        storage_context: StorageContext,
+        embed_model: EmbedType,
+        transformations: list[TransformComponent],
+        count_workers: int,
+        *args: Any,
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(storage_context, embed_model, transformations, *args, **kwargs)
+        self.count_workers = count_workers
+        assert (
+            len(self.transformations) >= 2
+        ), "Embeddings must be in the transformations"
+        assert count_workers > 0, "count_workers must be > 0"
+        self.count_workers = count_workers
+        # We are doing our own multiprocessing
+        # To do not collide with the multiprocessing of huggingface, we disable it
+        os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+        # doc_q stores parsed files as Document chunks.
+        # Using a shallow queue causes the filesystem parser to block
+        # when it reaches capacity. This ensures it doesn't outpace the
+        # computationally intensive embeddings phase, avoiding unnecessary
+        # memory consumption.  The semaphore is used to bound the async worker
+        # embedding computations to cause the doc Q to fill and block.
+        self.doc_semaphore = multiprocessing.Semaphore(
+            self.count_workers
+        )  # limit the doc queue to # items.
+        self.doc_q: Queue[tuple[str, str | None, list[Document] | None]] = Queue(20)
+        # node_q stores documents parsed into nodes (embeddings).
+        # Larger queue size so we don't block the embedding workers during a slow
+        # index update.
+        self.node_q: Queue[
+            tuple[str, str | None, list[Document] | None, list[BaseNode] | None]
+        ] = Queue(40)
+        threading.Thread(target=self._doc_to_node, daemon=True).start()
+        threading.Thread(target=self._write_nodes, daemon=True).start()
+
+    def _doc_to_node(self) -> None:
+        # Parse documents into nodes
+        with multiprocessing.pool.ThreadPool(processes=self.count_workers) as pool:
+            while True:
+                try:
+                    cmd, file_name, documents = self.doc_q.get(
+                        block=True
+                    )  # Documents for a file
+                    if cmd == "process":
+                        # Push CPU/GPU embedding work to the worker pool
+                        # Acquire semaphore to control access to worker pool
+                        self.doc_semaphore.acquire()
+                        pool.apply_async(
+                            self._doc_to_node_worker, (file_name, documents)
+                        )
+                    elif cmd == "quit":
+                        break
+                finally:
+                    if cmd != "process":
+                        self.doc_q.task_done()  # unblock Q joins
+
+    def _doc_to_node_worker(self, file_name: str, documents: list[Document]) -> None:
+        # CPU/GPU intensive work in its own process
+        try:
+            nodes = run_transformations(
+                documents,  # type: ignore[arg-type]
+                self.transformations,
+                show_progress=self.show_progress,
+            )
+            self.node_q.put(("process", file_name, documents, nodes))
+        finally:
+            self.doc_semaphore.release()
+            self.doc_q.task_done()  # unblock Q joins
+
+    def _save_docs(
+        self, files: list[str], documents: list[Document], nodes: list[BaseNode]
+    ) -> None:
+        try:
+            logger.info(
+                f"Saving {len(files)} files ({len(documents)} documents / {len(nodes)} nodes)"
+            )
+            self._index.insert_nodes(nodes)
+            for document in documents:
+                self._index.docstore.set_document_hash(
+                    document.get_doc_id(), document.hash
+                )
+            self._save_index()
+        except Exception:
+            # Tell the user so they can investigate these files
+            logger.exception(f"Processing files {files}")
+        finally:
+            # Clearing work, even on exception, maintains a clean state.
+            nodes.clear()
+            documents.clear()
+            files.clear()
+
+    def _write_nodes(self) -> None:
+        # Save nodes to index.  I/O intensive.
+        node_stack: list[BaseNode] = []
+        doc_stack: list[Document] = []
+        file_stack: list[str] = []
+        while True:
+            try:
+                cmd, file_name, documents, nodes = self.node_q.get(block=True)
+                if cmd in ("flush", "quit"):
+                    if file_stack:
+                        self._save_docs(file_stack, doc_stack, node_stack)
+                    if cmd == "quit":
+                        break
+                elif cmd == "process":
+                    node_stack.extend(nodes)  # type: ignore[arg-type]
+                    doc_stack.extend(documents)  # type: ignore[arg-type]
+                    file_stack.append(file_name)  # type: ignore[arg-type]
+                    # Constant saving is heavy on I/O - accumulate to a threshold
+                    if len(node_stack) >= self.NODE_FLUSH_COUNT:
+                        self._save_docs(file_stack, doc_stack, node_stack)
+            finally:
+                self.node_q.task_done()
+
+    def _flush(self) -> None:
+        self.doc_q.put(("flush", None, None))
+        self.doc_q.join()
+        self.node_q.put(("flush", None, None, None))
+        self.node_q.join()
+
+    def ingest(self, file_name: str, file_data: Path) -> list[Document]:
+        documents = IngestionHelper.transform_file_into_documents(file_name, file_data)
+        self.doc_q.put(("process", file_name, documents))
+        self._flush()
+        return documents
+
+    def bulk_ingest(self, files: list[tuple[str, Path]]) -> list[Document]:
+        docs = []
+        for file_name, file_data in eta(files):
+            try:
+                documents = IngestionHelper.transform_file_into_documents(
+                    file_name, file_data
+                )
+                self.doc_q.put(("process", file_name, documents))
+                docs.extend(documents)
+            except Exception:
+                logger.exception(f"Skipping {file_data.name}")
+        self._flush()
+        return docs
+
+
 def get_ingestion_component(
     storage_context: StorageContext,
     embed_model: EmbedType,
@@ -336,6 +502,13 @@ def get_ingestion_component(
             transformations=transformations,
             count_workers=settings.embedding.count_workers,
         )
+    elif ingest_mode == "pipeline":
+        return PipelineIngestComponent(
+            storage_context=storage_context,
+            embed_model=embed_model,
+            transformations=transformations,
+            count_workers=settings.embedding.count_workers,
+        )
     else:
         return SimpleIngestComponent(
             storage_context=storage_context,

private_gpt/settings/settings.py

@@ -155,13 +155,14 @@ class HuggingFaceSettings(BaseModel):
 
 class EmbeddingSettings(BaseModel):
     mode: Literal["huggingface", "openai", "azopenai", "sagemaker", "ollama", "mock"]
-    ingest_mode: Literal["simple", "batch", "parallel"] = Field(
+    ingest_mode: Literal["simple", "batch", "parallel", "pipeline"] = Field(
         "simple",
         description=(
             "The ingest mode to use for the embedding engine:\n"
             "If `simple` - ingest files sequentially and one by one. It is the historic behaviour.\n"
             "If `batch` - if multiple files, parse all the files in parallel, "
             "and send them in batch to the embedding model.\n"
+            "In `pipeline` - The Embedding engine is kept as busy as possible\n"
             "If `parallel` - parse the files in parallel using multiple cores, and embedd them in parallel.\n"
             "`parallel` is the fastest mode for local setup, as it parallelize IO RW in the index.\n"
             "For modes that leverage parallelization, you can specify the number of "
@@ -174,6 +175,7 @@ class EmbeddingSettings(BaseModel):
             "The number of workers to use for file ingestion.\n"
             "In `batch` mode, this is the number of workers used to parse the files.\n"
             "In `parallel` mode, this is the number of workers used to parse the files and embed them.\n"
+            "In `pipeline` mode, this is the number of workers that can perform embeddings.\n"
             "This is only used if `ingest_mode` is not `simple`.\n"
             "Do not go too high with this number, as it might cause memory issues. (especially in `parallel` mode)\n"
             "Do not set it higher than your number of threads of your CPU."

private_gpt/utils/eta.py

@@ -0,0 +1,122 @@
+import datetime
+import logging
+import math
+import time
+from collections import deque
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+
+def human_time(*args: Any, **kwargs: Any) -> str:
+    def timedelta_total_seconds(timedelta: datetime.timedelta) -> float:
+        return (
+            timedelta.microseconds
+            + 0.0
+            + (timedelta.seconds + timedelta.days * 24 * 3600) * 10**6
+        ) / 10**6
+
+    secs = float(timedelta_total_seconds(datetime.timedelta(*args, **kwargs)))
+    # We want (ms) precision below 2 seconds
+    if secs < 2:
+        return f"{secs * 1000}ms"
+    units = [("y", 86400 * 365), ("d", 86400), ("h", 3600), ("m", 60), ("s", 1)]
+    parts = []
+    for unit, mul in units:
+        if secs / mul >= 1 or mul == 1:
+            if mul > 1:
+                n = int(math.floor(secs / mul))
+                secs -= n * mul
+            else:
+                # >2s we drop the (ms) component.
+                n = int(secs)
+            if n:
+                parts.append(f"{n}{unit}")
+    return " ".join(parts)
+
+
+def eta(iterator: list[Any]) -> Any:
+    """Report an ETA after 30s and every 60s thereafter."""
+    total = len(iterator)
+    _eta = ETA(total)
+    _eta.needReport(30)
+    for processed, data in enumerate(iterator, start=1):
+        yield data
+        _eta.update(processed)
+        if _eta.needReport(60):
+            logger.info(f"{processed}/{total} - ETA {_eta.human_time()}")
+
+
+class ETA:
+    """Predict how long something will take to complete."""
+
+    def __init__(self, total: int):
+        self.total: int = total  # Total expected records.
+        self.rate: float = 0.0  # per second
+        self._timing_data: deque[tuple[float, int]] = deque(maxlen=100)
+        self.secondsLeft: float = 0.0
+        self.nexttime: float = 0.0
+
+    def human_time(self) -> str:
+        if self._calc():
+            return f"{human_time(seconds=self.secondsLeft)} @ {int(self.rate * 60)}/min"
+        return "(computing)"
+
+    def update(self, count: int) -> None:
+        # count should be in the range 0 to self.total
+        assert count > 0
+        assert count <= self.total
+        self._timing_data.append((time.time(), count))  # (X,Y) for pearson
+
+    def needReport(self, whenSecs: int) -> bool:
+        now = time.time()
+        if now > self.nexttime:
+            self.nexttime = now + whenSecs
+            return True
+        return False
+
+    def _calc(self) -> bool:
+        # A sample before a prediction.   Need two points to compute slope!
+        if len(self._timing_data) < 3:
+            return False
+
+        # http://en.wikipedia.org/wiki/Pearson_product-moment_correlation_coefficient
+        # Calculate means and standard deviations.
+        samples = len(self._timing_data)
+        # column wise sum of the timing tuples to compute their mean.
+        mean_x, mean_y = (
+            sum(i) / samples for i in zip(*self._timing_data, strict=False)
+        )
+        std_x = math.sqrt(
+            sum(pow(i[0] - mean_x, 2) for i in self._timing_data) / (samples - 1)
+        )
+        std_y = math.sqrt(
+            sum(pow(i[1] - mean_y, 2) for i in self._timing_data) / (samples - 1)
+        )
+
+        # Calculate coefficient.
+        sum_xy, sum_sq_v_x, sum_sq_v_y = 0.0, 0.0, 0
+        for x, y in self._timing_data:
+            x -= mean_x
+            y -= mean_y
+            sum_xy += x * y
+            sum_sq_v_x += pow(x, 2)
+            sum_sq_v_y += pow(y, 2)
+        pearson_r = sum_xy / math.sqrt(sum_sq_v_x * sum_sq_v_y)
+
+        # Calculate regression line.
+        # y = mx + b where m is the slope and b is the y-intercept.
+        m = self.rate = pearson_r * (std_y / std_x)
+        y = self.total
+        b = mean_y - m * mean_x
+        x = (y - b) / m
+
+        # Calculate fitted line (transformed/shifted regression line horizontally).
+        fitted_b = self._timing_data[-1][1] - (m * self._timing_data[-1][0])
+        fitted_x = (y - fitted_b) / m
+        _, count = self._timing_data[-1]  # adjust last data point progress count
+        adjusted_x = ((fitted_x - x) * (count / self.total)) + x
+        eta_epoch = adjusted_x
+
+        self.secondsLeft = max([eta_epoch - time.time(), 0])
+        return True

settings-local.yaml

@@ -1,3 +1,4 @@
+# poetry install --extras "ui llms-llama-cpp vector-stores-qdrant embeddings-huggingface"
 server:
   env_name: ${APP_ENV:local}