Merge upstream/main into docs/ocr-v2-defaults

Resolve multimodal-extraction.md conflicts: keep chart pdfium-path
callout from main and PR OCR v2 deployment breakdown.

Author: kheiss-uwzoo

.github/workflows/reusable-pypi-publish.yml

@@ -84,7 +84,8 @@ jobs:
           DIST_FILES=(./dist/nemo_retriever/dist/*.whl ./dist/nemo_retriever/dist/*.tar.gz)
           echo "After legacy glob: count=${#DIST_FILES[@]} files=${DIST_FILES[*]:-<none>}"
 
-          # Flat layout (older upload-artifact glob uploads): ./dist/*
+          # Flat layout (upload-artifact strips to the common parent, so wheels
+          # often land directly under ./dist/): ./dist/*
           if [ "${#DIST_FILES[@]}" -eq 0 ]; then
             DIST_FILES=(./dist/*.whl ./dist/*.tar.gz)
             echo "After flat glob: count=${#DIST_FILES[@]} files=${DIST_FILES[*]:-<none>}"

Dockerfile

@@ -4,6 +4,8 @@
 # syntax=docker/dockerfile:1.3
 #
 # Build from repo root: docker build -f Dockerfile -t nemo-retriever .
+# Service (NIM-forwarding): docker build -f Dockerfile --target service -t nemo-retriever-service .
+# Service + in-pod HF:      docker build -f Dockerfile --target service-gpu -t nemo-retriever-service-gpu .
 # Runtime ffmpeg/ffprobe install for service image: docker run -e INSTALL_FFMPEG=true nemo-retriever-service
 # Run: docker run nemo-retriever  (shell with venv active)
 # Run with dev mount: docker run -v $(pwd):/workspace -it nemo-retriever   (code changes reflect without rebuild)
@@ -116,6 +118,22 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     . /opt/retriever_runtime/bin/activate \
     && uv pip install -e "./nemo_retriever[service]"
 
+# GPU service install: in-pod Hugging Face models + multimedia (ASR, SVG).
+# Build target: service-gpu
+FROM base AS install-service-gpu
+
+WORKDIR /workspace
+
+ENV PYTHONUNBUFFERED=1
+ENV VIRTUAL_ENV=/opt/retriever_runtime
+ENV PATH=/opt/retriever_runtime/bin:$PATH
+
+SHELL ["/bin/bash", "-c"]
+RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=cache,target=/root/.cache/uv \
+    . /opt/retriever_runtime/bin/activate \
+    && uv pip install -e "./nemo_retriever[service,local,multimedia]"
+
 # Default: run in-process pipeline (help if no args)
 CMD ["/bin/bash"]
 
@@ -167,3 +185,45 @@ USER nemo
 ENTRYPOINT ["/usr/local/bin/retriever-service-entrypoint"]
 
 CMD ["retriever", "service", "start", "--config", "/etc/nemo-retriever/retriever-service.yaml"]
+
+# ---------------------------------------------------------------------------
+# GPU service profile: FastAPI ingest service with in-pod Hugging Face models.
+#
+# Build:  docker build -f Dockerfile --target service-gpu \
+#             -t nemo-retriever-service-gpu .
+#
+# Run (requires --gpus all and local_models.enabled in config):
+#   docker run --rm --gpus all -p 7670:7670 \
+#     -v /host/retriever-service.yaml:/etc/nemo-retriever/retriever-service.yaml:ro \
+#     nemo-retriever-service-gpu
+# ---------------------------------------------------------------------------
+FROM install-service-gpu AS service-gpu
+
+ENV NEMO_RETRIEVER_SERVICE_CONFIG=/etc/nemo-retriever/retriever-service.yaml
+
+ENV PATH=/opt/retriever_runtime/bin:$PATH
+
+COPY docker/scripts/retriever_service_entrypoint.sh /usr/local/bin/retriever-service-entrypoint
+COPY docker/scripts/retriever_install_ffmpeg.sh /usr/local/sbin/retriever-install-ffmpeg
+
+RUN chmod a+rx /usr/local/bin/uv /usr/local/bin/uvx /usr/local/bin/retriever-service-entrypoint \
+        /usr/local/sbin/retriever-install-ffmpeg \
+    && chmod -R a+rX /opt/uv \
+    && groupadd -r -g 1000 nemo && useradd -r -u 1000 -g nemo -d /workspace -s /sbin/nologin nemo \
+    && printf '%s\n' \
+         'nemo ALL=(root) NOPASSWD: /usr/local/sbin/retriever-install-ffmpeg' \
+         > /etc/sudoers.d/nemo-ffmpeg \
+    && chmod 0440 /etc/sudoers.d/nemo-ffmpeg \
+    && visudo -cf /etc/sudoers.d/nemo-ffmpeg \
+    && mkdir -p /etc/nemo-retriever /var/lib/nemo-retriever \
+    && cp /workspace/nemo_retriever/src/nemo_retriever/service/retriever-service.yaml \
+            "${NEMO_RETRIEVER_SERVICE_CONFIG}" \
+    && chown -R nemo:nemo /workspace /etc/nemo-retriever /var/lib/nemo-retriever /opt/retriever_runtime
+
+EXPOSE 7670
+
+USER nemo
+
+ENTRYPOINT ["/usr/local/bin/retriever-service-entrypoint"]
+
+CMD ["retriever", "service", "start", "--config", "/etc/nemo-retriever/retriever-service.yaml"]

README.md

@@ -36,10 +36,10 @@ For small-scale workloads, such as workloads of fewer than 100 PDFs, you can use
 After [following the quickstart installation steps](nemo_retriever), you can start ingesting content like with the following snippet:
 ```python
 from nemo_retriever import create_ingestor
-from nemo_retriever.io import to_markdown, to_markdown_by_page
+from nemo_retriever.common.io import to_markdown, to_markdown_by_page
 from pathlib import Path
 
-documents = [str(Path("../data/multimodal_test.pdf"))]
+documents = [str(Path("data/multimodal_test.pdf"))]
 ingestor = create_ingestor(run_mode="batch")
 
 # ingestion tasks are chainable and defined lazily
@@ -93,7 +93,7 @@ dict_keys([1, 2, 3])
 To query for relevant snippets of the ingested content, and use them with an LLM to generate answers, use the following code.
 
 ```python
-from nemo_retriever.retriever import Retriever
+from nemo_retriever.graph.retriever import Retriever
 from openai import OpenAI
 import os
 

docs/docs/extraction/audio-video.md

@@ -100,6 +100,8 @@ Instead of running the pipeline locally, you can call Parakeet through [build.nv
 
 2. Run inference from Python with the hosted gRPC endpoint and credentials from that page (the example below uses the default hosted gRPC hostname; confirm values in the **Get API Key** flow for your deployment). Pass hosted endpoint, function ID, and API key through `ASRParams` (`audio_endpoints`, `function_id`, `auth_token`).
 
+    For parameter details, refer to the [Python API guide](nemo-retriever-api-reference.md).
+
     ```python
     from nemo_retriever import create_ingestor
     from nemo_retriever.common.params.models import ASRParams

docs/docs/extraction/concepts.md

@@ -2,40 +2,40 @@
 
 These terms appear throughout NeMo Retriever Library documentation.
 
-## Job
+## Job { #job }
 
-An **ingestion job** is a unit of work you run on input content (documents, audio, video, and other supported types). You submit jobs through the **ingestor Python API** (for example `Ingestor` task chains such as `.extract(...)`) or the **`retriever ingest` CLI**—not by posting a standalone JSON job document. Default tasks target strong recall; customize behavior with task keyword arguments (including chunking and splitting on `.extract()`) or custom UDF-style operations ([NeMo Retriever graph](https://github.com/NVIDIA/NeMo-Retriever/tree/main/nemo_retriever/src/nemo_retriever/graph#nemo-retriever-graph)). Results are structured metadata and annotations (Ray Dataset, pandas `DataFrame`, or similar).
+An **ingestion job** is a unit of work you run on input content (documents, audio, video, and other supported types). You submit jobs through the **ingestor Python API** (for example `Ingestor` task chains such as `.extract(...)`) or the **`retriever ingest` CLI**—not by posting a standalone JSON job document. Default tasks target strong recall; customize behavior with task keyword arguments (including chunking and splitting on `.extract()`) or custom UDF-style operations. For UDFs and other extension paths, refer to [Customize & extend](customize-extend.md). Results are structured metadata and annotations (Ray Dataset, pandas `DataFrame`, or similar).
 
-## Pipeline and tasks
+## Pipeline and tasks { #pipeline-and-tasks }
 
-NeMo Retriever Library does **not** run one static pipeline on every document. You configure **tasks** such as parsing, chunking, embedding, storage, and filtering per job. Related topics: [Extending/Customizing NeMo Retriever Library with custom code](https://github.com/NVIDIA/NeMo-Retriever/tree/main/nemo_retriever/src/nemo_retriever/graph#nemo-retriever-graph).
+NeMo Retriever Library does **not** run one static pipeline on every document. You configure **tasks** such as parsing, chunking, embedding, storage, and filtering per job. For UDFs, custom graph stages, and other extension paths, refer to [Customize & extend](customize-extend.md).
 
-## Extraction metadata
+## Extraction metadata { #extraction-metadata }
 
 Output is a **Ray Dataset** (Ray Data) or **pandas** `DataFrame` listing extracted objects (text regions, tables, images, and so on), processing notes, and timing or trace data. Field-level detail is in the [metadata reference](content-metadata.md).
 
-## Embeddings and retrieval
+## Embeddings and retrieval { #embeddings-and-retrieval }
 
-Optionally, the library can compute **embeddings** for extracted content and store vectors in [LanceDB](https://lancedb.com/) for downstream semantic search in your application. For upload and retrieval APIs, see [Vector databases](vdbs.md). For multimodal (VLM) embedding options, see [Multimodal embeddings (VLM)](embedding.md).
+Optionally, the library can compute **embeddings** for extracted content and store vectors in [LanceDB](https://lancedb.com/) for downstream semantic search in your application. For upload and retrieval APIs, refer to [Vector databases](vdbs.md). For multimodal (VLM) embedding options, refer to [Multimodal embeddings (VLM)](embedding.md).
 
 ## Chunking { #chunking }
 
 Chunking is built into the `.extract()` task and depends on **content type**:
 
 - **PDF, DOCX, and PPTX** — Text is grouped using built-in **page** boundaries (one chunk per page where the format has pages).
-- **Plain text (`.txt`) and HTML** — Formats without natural page breaks are split into segments of **1024 tokens** by default, using the [Llama 3.2 1B tokenizer](https://huggingface.co/meta-llama/Llama-3.2-1B) so chunk boundaries stay aligned with the default embedding tokenizer. The NeMo Retriever container image bundles this tokenizer, so default text chunking does not require a Hugging Face access token. See [Token-based splitting](#token-based-splitting) and [Environment variables](environment-config.md) for overrides and other runtimes.
-- **Audio and video** — Media is split into **segments** for decoding and ASR using ffmpeg-based rules (configurable **size**, **time**, or **frame** split modes in the media chunking stage). With the Parakeet ASR path, you can optionally emit **sentence-like segments** using `extract_audio_params={"segment_audio": True}`; see [Speech and audio extraction](audio-video.md#speech-and-audio-extraction).
+- **Plain text (`.txt`) and HTML** — Formats without natural page breaks are split into segments of **1024 tokens** by default, using the [Llama 3.2 1B tokenizer](https://huggingface.co/meta-llama/Llama-3.2-1B) so chunk boundaries stay aligned with the default embedding tokenizer. The NeMo Retriever container image bundles this tokenizer, so default text chunking does not require a Hugging Face access token. Refer to [Token-based splitting](#token-based-splitting) and [Environment variables](environment-config.md) for overrides and other runtimes.
+- **Audio and video** — Media is split into **segments** for decoding and ASR using ffmpeg-based rules (configurable **size**, **time**, or **frame** split modes in the media chunking stage). With the Parakeet ASR path, you can optionally emit **sentence-like segments** using `extract_audio_params={"segment_audio": True}`; refer to [Speech and audio extraction](audio-video.md#speech-and-audio-extraction).
 
-For PDF parallelism before Ray processing (large files), see [PDF pre-splitting for parallel ingest](nemo-retriever-api-reference.md#pdf-pre-splitting-for-parallel-ingest).
+For PDF parallelism before Ray processing (large files), refer to [PDF pre-splitting for parallel ingest](nemo-retriever-api-reference.md#pdf-pre-splitting-for-parallel-ingest).
 
 ### Token-based splitting { #token-based-splitting }
 
-Token-based splitting uses the Llama 3.2 1B tokenizer (default `meta-llama/Llama-3.2-1B`) with configurable `max_tokens` and `overlap_tokens` when you add an explicit `.split(...)` stage or when the pipeline applies the default text segmentation for unstructured text. In the shipped NeMo Retriever container, tokenizer assets are included locally, so you do not need `HF_ACCESS_TOKEN` for this default path. If your runtime loads the tokenizer from the Hugging Face Hub instead (for example, some library-only installs), set `HF_ACCESS_TOKEN` or pass `hf_access_token` in task params when the Hub requires it. Details appear in the [Python API guide](nemo-retriever-api-reference.md).
+Token-based splitting uses the Llama 3.2 1B tokenizer (default `meta-llama/Llama-3.2-1B`) with configurable `max_tokens` and `overlap_tokens` when you add an explicit `.split(...)` stage or when the pipeline applies the default text segmentation for unstructured text. In the shipped NeMo Retriever container, tokenizer assets are included locally, so you do not need `HF_ACCESS_TOKEN` for this default path. If your runtime loads the tokenizer from the Hugging Face Hub instead (for example, some library-only installs), set `HF_ACCESS_TOKEN` or pass `hf_access_token` in task params when the Hub requires it. For parameter details, refer to the [Python API guide](nemo-retriever-api-reference.md).
 
-## Deployment modes
+## Deployment modes { #deployment-modes }
 
-- **Library mode** — Run without the full container stack where appropriate; see [Deployment options](deployment-options.md).
-- **Kubernetes / Helm (self-hosted)** — See [Deploy (Helm chart)](https://github.com/NVIDIA/NeMo-Retriever/blob/main/nemo_retriever/helm/README.md) and [deployment options](deployment-options.md) for running the full microservices pipeline on your infrastructure.
+- **Library mode** — Run without the full container stack where appropriate; refer to [Deployment options](deployment-options.md).
+- **Kubernetes / Helm (self-hosted)** — Refer to [Deploy (Helm chart)](https://github.com/NVIDIA/NeMo-Retriever/blob/main/nemo_retriever/helm/README.md) and [deployment options](deployment-options.md) for running the full microservices pipeline on your infrastructure.
 - **Notebooks** — [Jupyter examples](starter-kits.md) for experimentation and RAG demos.
 
 For a concise comparison, refer to [Deployment options](deployment-options.md).

docs/docs/extraction/customize-extend.md

@@ -0,0 +1,66 @@
+# Customize & extend
+
+NeMo Retriever Library ships with defaults tuned for strong recall on common document types. When those defaults are not enough, you can extend the library at several levels—from task keyword arguments on the fluent ingestor API through custom graph operators and vector-database adapters.
+
+Use this page to choose an extension path and find the detailed guides in the repository.
+
+The following table maps common needs to the right section:
+
+| If you need to… | Start here |
+|-----------------|------------|
+| Tune extraction, chunking, embedding, or upload without new code | [Start with task configuration](#start-with-task-configuration) |
+| Add a small Python transformation between pipeline stages | [User-defined functions (UDFs)](#user-defined-functions-udfs) |
+| Build or reuse operators stage-by-stage | [Custom graph pipelines](#custom-graph-pipelines) |
+| Store vectors in a backend other than LanceDB | [Custom vector databases](#custom-vector-databases) |
+
+## On this page { #on-this-page }
+
+- [Start with task configuration](#start-with-task-configuration)
+- [User-defined functions (UDFs)](#user-defined-functions-udfs)
+- [Custom graph pipelines](#custom-graph-pipelines)
+- [Custom vector databases](#custom-vector-databases)
+- [Related Topics](#related-topics)
+
+## Start with task configuration { #start-with-task-configuration }
+
+Most customization does not require new code. Chain tasks on `create_ingestor(...)` and pass keyword arguments to control extraction, chunking, embedding, and storage—for example `extract_method`, chunking and splitting options on `.extract()`, `embed_modality` on `.embed()`, and `vdb_op` / `vdb_kwargs` on `.vdb_upload()`.
+
+For parameter details, refer to the [Python API guide](nemo-retriever-api-reference.md). For chunking behavior and pipeline concepts, refer to [Concepts](concepts.md).
+
+## User-defined functions (UDFs) { #user-defined-functions-udfs }
+
+A **user-defined function (UDF)** wraps your Python logic as a first-class pipeline stage. In the graph model, `UDFOperator` turns a plain callable into an operator you can chain with built-in stages—for example to normalize HTML, apply a custom split, or call an external service between extract and embed steps.
+
+Use UDFs when you need a small, self-contained transformation that is not covered by task keyword arguments.
+
+### Repository guides
+
+- [NeMo Retriever graph README — `UDFOperator`](https://github.com/NVIDIA/NeMo-Retriever/tree/main/nemo_retriever/src/nemo_retriever/graph#using-udfoperator) — API, lifecycle, and when to use `UDFOperator` versus a custom operator class
+- [NimClient and custom NIM endpoints](https://github.com/NVIDIA/NeMo-Retriever/blob/main/nemo_retriever/developer_docs/nimclient.md#nimclient-and-custom-nim-endpoints) — call custom or self-hosted NIM microservices from UDF stages
+
+## Custom graph pipelines { #custom-graph-pipelines }
+
+When you need to compose pipelines stage-by-stage, reuse operators across workflows, or run the same graph in-process or with Ray Data, use the **graph execution model** instead of (or alongside) the fluent `GraphIngestor` API.
+
+The graph package provides `AbstractOperator`, executors (`InprocessExecutor`, `RayDataExecutor`), and operator chaining with `>>`. Built-in ingestion operators live under `nemo_retriever.operators`; you can add your own operators or UDF stages anywhere in the chain.
+
+For the full guide—including custom operator classes, executors, and graph shape constraints—refer to the [NeMo Retriever graph README](https://github.com/NVIDIA/NeMo-Retriever/tree/main/nemo_retriever/src/nemo_retriever/graph#nemo-retriever-graph).
+
+## Custom vector databases { #custom-vector-databases }
+
+The supported user path for vector storage is **[LanceDB](vdbs.md)** (`vdb_op="lancedb"`). That page covers upload, semantic retrieval, metadata filtering, and LanceDB deployment characteristics.
+
+To integrate a different vector store, implement the [`VDB`](https://github.com/NVIDIA/NeMo-Retriever/blob/main/nemo_retriever/src/nemo_retriever/common/vdb/adt_vdb.py) interface and wire it through graph [`IngestVdbOperator`](https://github.com/NVIDIA/NeMo-Retriever/blob/main/nemo_retriever/src/nemo_retriever/operators/vdb.py) / [`RetrieveVdbOperator`](https://github.com/NVIDIA/NeMo-Retriever/blob/main/nemo_retriever/src/nemo_retriever/operators/vdb.py). NVIDIA validates the first-party LanceDB operator; you are responsible for testing and maintaining other backends.
+
+### Repository guides
+
+- [Vector DB package (source)](https://github.com/NVIDIA/NeMo-Retriever/tree/main/nemo_retriever/src/nemo_retriever/common/vdb) — `VDB` abstract base and LanceDB reference implementation
+
+Partner and blueprint integrations (Elasticsearch, Pinecone, Teradata, and others) are summarized on [Vector databases — Vector database partners](vdbs.md#vector-database-partners).
+
+## Related Topics { #related-topics }
+
+- [Concepts — Pipeline and tasks](concepts.md#pipeline-and-tasks)
+- [Vector databases](vdbs.md)
+- [Multimodal embeddings (VLM)](embedding.md)
+- [Python API guide](nemo-retriever-api-reference.md)