Document ingestion splits off as agent infrastructure, and fidelity becomes the metric

Document ingestion as agent infrastructure is already a well-funded category. Reducto, LlamaParse, Unstructured, Tensorlake, Docling and the hyperscalers all compete in it. Both product ideas are partly taken. On the parser: Reducto already outputs strikethrough and underline markup as change-tracking cues. Tensorlake already keeps DOCX insertions, deletions and comments as semantic tags, and it targets legal tech. Several open-source libraries do the same for DOCX. What none of them seem to offer is a typed annotation schema built for compliance work. That would mean showing who made a revision and when, the status of each revision, emphasis that carries meaning (such as "must" in bold), superseded clauses, and version lineage across policy documents, all handled the same way across scanned PDFs and DOCX. It would also need audit-trail guarantees that bank model-risk and compliance teams require. That is a narrow vertical layer, and incumbents could add it as a feature, so it is probably best sold as a compliance-specific layer on top of an existing parser rather than as a competing parser. On the benchmark: ParseBench from LlamaIndex already claims to be the agent-focused parsing benchmark and already scores semantic formatting. OmniDocBench and RD-TableBench cover tables. However, every serious benchmark is either run by a vendor or academic and centred on PDFs. None we found tests whether DOCX revision marks and deletions survive parsing, or whether an agent's downstream decisions change when those cues are lost. That is the cue-loss task design. A neutral third party offering private audits on a buyer's own legal or insurance documents is the clearest opening, because vendor-run leaderboards have an obvious conflict of interest. The risk is that this becomes a small consulting business rather than a product, and that vendors will simply publish their own competing benchmarks.

• Reducto

  A well-funded ingestion API that sells itself as the document layer for LLMs. It already offers a change-tracking option that wraps struck-through text in s tags and underlined text in u tags. For digital PDFs and DOCX it reads embedded formatting, and for scans it uses a vision model. It also publishes its own benchmarks, including RD-TableBench.

• LlamaIndex ParseBench

  A public benchmark of about 2,000 enterprise pages, released in April 2026 and billed as a document parsing benchmark for AI agents. Semantic formatting is one of its five scoring dimensions, and its authors say no earlier benchmark scored formatting. It is vendor-run, and LlamaParse comes out on top.

• LlamaParse (LlamaCloud)

  A hosted parser with agentic tiers aimed at agent and RAG pipelines. LlamaIndex also publishes material on strikethrough detection.

• Tensorlake

  Parses DOCX tracked changes into ins and del tags and keeps comments as anchored spans, along with bounding boxes and page numbers. It markets this directly to legal tech, which is close to the policy-grade parser idea.

• OmniDocBench

  An open CVPR 2025 benchmark from Shanghai AI Lab. It scores text, tables (using TEDS), formulas and layout across diverse PDFs, and has a public leaderboard.

• Adeu

  An open-source MCP server and SDK that turns DOCX into token-efficient Markdown so agents can redline without breaking formatting.

• Redlines (Elixir)

  An open-source library that pulls tracked insertions and deletions out of DOCX and formats them for LLM prompts.

• docx-review-core

  A Rust crate that turns DOCX into a normalized document model with blocks, tracked changes and comments.

• Unstructured, Docling, Azure Document Intelligence, Google Document AI, AWS Textract

  General ingestion and IDP options that buyers compare against. Azure has no documented strikethrough output, so a gap is left at the hyperscaler level.

Checked 2026-10-22