Add inline text ingestion support

[jioffe502]

Summary

• add .texts() for ingesting raw Python strings without temporary files
• infer each source modality automatically, including inline://... text sources
• allow inline text alongside files and buffers through the existing modality planner
• keep chunking configuration on the established .extract(split_config=...) interface; no separate .extract_txt() or Graph .split() API
• reuse the normal split, embed, VDB, and service paths in in-process, batch, and service modes

Closes #2345.

Validation

• full local unit suite: 2,685 passed, 131 skipped, 12 deselected
• focused ingestion, manifest, splitter, actor, and service suite: 307 passed, 2 skipped, 4 deselected
• 2 local-Ray file/inline parity and mixed-source tests passed
• service worker test passed with default auto-routing and preserved inline source identity
• Black, Flake8, and diff whitespace checks passed on the changed files

[greptile-apps]

Greptile Summary

This PR adds a .texts() method to both GraphIngestor and ServiceIngestor enabling raw Python strings to be ingested without temporary files. Inline sources receive deterministic inline://NNNNNNNN identifiers and route through the existing text-split, embed, VDB, and service paths without any new graph stages.

• New primitives in common/inline_text.py give the feature a clean, tested foundation; text_to_chunks_df is extracted from the file/bytes adapters and shared across all three paths.
• Routing integration: _classified_input_paths, build_input_manifest, and infer_extraction_mode_from_filename all classify inline:// URIs as \"txt\" so the manifest planner and service router handle inline sources automatically alongside files.
• Service mode: inline texts are packed as InMemoryUpload with a classification_filename header so the server-side auto-router can identify and process them without touching the filesystem; blank-corpus and empty-list early-exits prevent unnecessary HTTP round-trips.

Confidence Score: 4/5

Safe to merge with one caveat: extract_txt() was a documented public method and is removed outright without a deprecation wrapper; existing callers will get an AttributeError at runtime.

The inline text feature itself is well-implemented: correct routing through manifests, no temp-file leakage, blank-corpus short-circuits, proper logging on exceptions, and broad test coverage across inprocess/batch/service modes. The only unresolved concern is the hard removal of the public GraphIngestor.extract_txt() method without a deprecation cycle.

Files Needing Attention: nemo_retriever/src/nemo_retriever/ingestor/graph_ingestor.py — the removal of extract_txt() is the single change that needs a second look before merging.

Important Files Changed

Filename	Overview
nemo_retriever/src/nemo_retriever/common/inline_text.py	New module introducing inline text primitives: normalize/validate, deterministic source IDs, blank-corpus guard, and inline URI detection.
nemo_retriever/src/nemo_retriever/common/modality/txt/split.py	Refactored to extract text_to_chunks_df from txt_file_to_chunks_df and txt_bytes_to_chunks_df, adding empty_text_chunks_df() and inline-source-path preservation in txt_bytes_to_chunks_df.
nemo_retriever/src/nemo_retriever/ingestor/graph_ingestor.py	Adds .texts(), removes .extract_txt() (breaking API change flagged in previous review), changes _extraction_mode default from pdf to None, and adds inline-text routing for inprocess/batch/branch paths.
nemo_retriever/src/nemo_retriever/ingestor/branch_extraction.py	Adds inline_rows field to ExtractionBranchExecutor; _partition_branch_inputs and _inline_rows_by_path helpers separate file, buffer, and inline inputs per branch for both inprocess and Ray batch paths.
nemo_retriever/src/nemo_retriever/service/service_ingestor.py	Adds .texts() to ServiceIngestor; _collect_inputs packs inline texts as InMemoryUpload items without writing temp files; blank-corpus early exit and _has_mixed_inline_sources auto-routing for mixed payloads.
nemo_retriever/src/nemo_retriever/service/client.py	Adds InMemoryUpload NamedTuple and UploadInput union type; extracts _upload_source helper; broadens ingest_documents and aingest_documents_stream to accept list[UploadInput] instead of list[Path].
nemo_retriever/src/nemo_retriever/operators/extract/txt/ray_data.py	TxtSplitCPUActor now dispatches on text column (inline path) vs bytes column (file/buffer path); exception handler now logs with exc_info=True via logger.warning before continuing.
nemo_retriever/src/nemo_retriever/service/utils/file_type.py	Adds early-return in infer_extraction_mode_from_filename for inline:// URIs so the server correctly routes in-memory text uploads to the text extractor.

_{Reviews (19): Last reviewed commit: "Remove redundant shutdown lease assertio..." | Re-trigger Greptile}

[jperez999]

why did you make this public? I think it should stay private no?

[jperez999]

is there anyway to refactor this internal logic and that from the aingest_document_stream internal into a single helper, that maybe calls a another function that handles the differences?

[jperez999]

what is the difference between this and regular empty_text_chunks_df? Seems confusing to have both. Is this only because of the columns available, and the dtype conversion?

[jperez999]

Like here, why can't we use just empty dataframe.