AgentContext

Document parsing that never loses the plot — or the page number.

AgentContext converts documents into clean Markdown and structured JSON, and unlike other converters, every block of output carries provenance: source, page, hierarchical section path, and character span. When your agent cites something, you can prove where it came from.

PDF / DOCX / HTML / Markdown  →  Markdown + JSON, fully traceable

Why another parser?

Tools like MarkItDown and Docling produce good Markdown — and then throw away most of what an AI agent needs to be trustworthy:

If your LLM answer says "revenue grew 12%", AgentContext lets you point at page 7, section "3. Financials > 3.2 Revenue" — automatically. (Bounding boxes land in v0.2 with layout analysis.)

Install

# core: txt / md / html parsing, zero dependencies
pip install agentcontext-core

# with PDF + DOCX support
pip install "agentcontext-core[pdf,docx]"

(The PyPI name is agentcontext-core — plain agentcontext is name-blocked by an unrelated existing project. The import is still import agentcontext.)

No GPU. No torch. No API keys. Pure parsing.

Quickstart

from agentcontext import Document

doc = Document.parse("report.pdf")

print(doc.to_markdown())              # clean, structured markdown
print(doc.to_json())                  # full document model, lossless

for block in doc.blocks:
    print(block.text[:60], "→ page", block.provenance.page)

for table in doc.tables:
    print(table.to_rows())            # structured cells, with provenance

Or from the command line:

agentcontext parse report.pdf                # writes report.md next to the source
agentcontext parse report.pdf --json         # writes report.json (full document model)
agentcontext parse report.pdf --cite inline  # markdown with provenance anchors

What the output looks like

--cite inline gives you Markdown that renders normally but carries its receipts:

# Refund Policy <!-- src: policy.md | Refund Policy -->

Customers may request a full refund within 30 days
of purchase. <!-- src: policy.md | Refund Policy -->

## Exceptions <!-- src: policy.md | Refund Policy > Exceptions -->

Digital goods are excluded. <!-- src: policy.md | Refund Policy > Exceptions -->

--json gives you the full Unified Document Model. Unknown provenance fields are explicit null, never omitted — a block without provenance is a bug:

{
  "udm_version": "0.1",
  "metadata": {
    "title": null, "author": null, "created": null,
    "source_path": "/abs/path/report.pdf",
    "sha256": "444cd23e4ba2b0a1…",
    "parser": "pdf", "parser_version": "pdf-parser/0.1"
  },
  "blocks": [
    {
      "type": "paragraph",
      "text": "Revenue grew 12% year over year...",
      "level": null,
      "provenance": {
        "source": "report.pdf",
        "page": 7,
        "section_path": "3. Financials > 3.2 Revenue",
        "bbox": null,
        "char_span": null,
        "confidence": 0.9,
        "parser": "pdf",
        "version": "pdf-parser/0.1"
      }
    }
  ],
  "tables": [ ... ]
}

Supported formats (v0.1)

• PDF (digital / text-layer)
• DOCX
• HTML
• Markdown (normalization + provenance) and plain text

OCR for scanned documents, PPTX, and XLSX are next on the roadmap.

Benchmarks

A public benchmark against MarkItDown and Docling on a golden corpus (papers, reports, contracts, invoices) — measuring text accuracy, structure accuracy, table cell accuracy, and provenance accuracy — is under construction: see BENCHMARKS.md.

We will publish the numbers even where we lose. Trust is the product.

Roadmap

• v0.1 (now): PDF/DOCX/HTML/MD → Markdown + JSON with full provenance. CLI + Python SDK.
• v0.2: OCR for scanned documents, PPTX/XLSX parsers, provenance-preserving chunking.
• v0.3: Embedding adapters, citation-aware retrieval helpers.
• Later: Context packages for agents — retrieval that returns not just chunks, but summaries, tables, entities, and citations in one structured payload.

The long-term vision is a full open context-engineering layer for AI agents (a working preview of the whole pipeline lives on the platform branch). The short-term promise is simpler: the most trustworthy parser you can put in a RAG pipeline.

Design principles

1. Provenance is not optional. A block without a source location is a bug.
2. Small core, pluggable edges. Parsers, OCR engines, and exporters implement a small protocol.
3. No heavyweight dependencies in core. pip install and go.
4. Honest benchmarks. Measured in CI, published publicly.

Contributing

The Parser protocol makes new formats easy to add:

from agentcontext import Document, Parser, register_parser

class EpubParser(Parser):
    name = "epub"
    version = "epub-parser/0.1"
    extensions = ("epub",)

    def parse(self, path: str) -> Document:
        ...

register_parser(EpubParser())

See CONTRIBUTING.md.

Author

Built by Harish — @harish-ai-engineer

License

Apache-2.0