Docling vs Google Cloud Document AI

Docling 60

Self-hosted rule. Free MIT software with nothing to buy, so 20 + 20 + 20, and no payment protocol (0). 60.

Google Cloud Document AI 20

No x402, MPP or L402 (0). Per-page prices for every processor published without a login (20). The first 1,000 OCR pages show at $0.00 and new customers get $300 of credit, but the Free Trial page says sign-up needs a credit card or other payment method (0). A person creates the project and billing account in a browser (0).

Docling 65

MIT, an OSI licence, with model licences noted as separate (30). Runs locally, remote services off by default and air-gapped use documented; no privacy statement is needed for local use, but model downloads from Hugging Face on first run are only partly described (22 of 30). Breaking changes are marked in the changelog, with no deprecation policy or dated notices (10 of 20). No telemetry found in the source; nothing to opt out of (15 of 20). 30+22+10+15 = 77.

Google Cloud Document AI 90

Closed service under the Google Cloud terms, with Apache-2.0 client libraries (15 of 30). The security page says online requests are processed in memory, batch documents are deleted after processing with a failsafe of one day, and content is never used to train Document AI models. The service terms' training restriction and the Data Processing Addendum agree. Layout Parser versions on Gemini 3.0 route globally, which the docs state (27 of 30). The lifecycle page promises six months' notice for stable versions and lists a deprecation date per version, and the Cloud terms promise 12 months before a backwards-incompatible API change. The legacy processor notice of 17 February 2026 gave until 30 June 2026, about four and a half months (18 of 20). us and eu multi-regions, seven single regions and a public sub-processor list (20).

Docling 53

Framework reading. No credentials for local use; docling-serve has one optional shared key, off by default (20 of 30). Remote services are off unless explicitly enabled, local file fetch is guarded by flags, and MCP tools carry read-only and destructive annotations; manipulation tools can delete items and save files (12 of 20). Converted documents are untrusted text and no prompt-injection guidance was found (3 of 15). Logging only, no audit trail (3 of 15). A disclosure policy with private reporting, 28 advisories published in public with fixed versions, under the LF AI & Data Foundation; no bug bounty (15 of 20). 20+12+3+3+15 = 53. The advisories themselves are the deduction.

Google Cloud Document AI 70

OAuth 2.0 bearer tokens from service accounts with IAM. The docs and samples send the token in the Authorization header, but Google's OAuth 2.0 guide documents sending it as an access_token query parameter, the v1 Discovery document (revision 20261026) lists access_token, oauth_token and key as query parameters, and on 1 November documentai.googleapis.com and us-documentai.googleapis.com answered a request with no token "missing required authentication credential" and one with a made-up ?access_token= "invalid authentication credentials". So a token can travel in the URL, the 10-point deduction other Google APIs take (20). roles/documentai.apiUser allows processing only, roles can be granted on one processor, and deny policies and VPC Service Controls are supported. Nothing asks for confirmation before a processor is deleted (17 of 20). The service returns text from untrusted documents and no guidance on injected instructions was found in the pages read (0 of 15). The audit logging page (last updated 27 October) lists each method by permission type. ProcessDocument writes Data Access logs, which Google Cloud leaves off until the owner enables them (13 of 15). google.com security.txt valid to 1 April 2030 with the reward programme. The Document AI security page (last updated 26 October) states ISO 27001, SOC 2 and SOC 3 audits, FedRAMP High and HIPAA. None of the Google Cloud security bulletins, from GCP-2019-001 to GCP-2026-074 (28 October 2026), names Document AI (20).

Docling 89

v2.148.0 on 5 November 2026, the day before this check (30). 13 releases from 9 October to 5 November (20). Open issues and pull requests are numerous, an AI triage workflow and PR reminders run, and six maintainers are named; replies not sampled (17 of 25). docling-mcp publishes to the official MCP registry on release under io.github.docling-project, per its workflow and server.json; the registry entry itself was not read (12 of 15). CI, a dependency-update workflow and uv.lock (10). 30+20+17+12+10 = 89.

Google Cloud Document AI 80

Read as a closed service with official SDKs. The newest release note is 21 September 2026, and the Python client 3.17.0 is dated 29 October 2026 in its changelog (30). Release notes on 21 September, Node.js releases dated 4 August, 8 September, 28 September and 8 October, and Python 3.16.0 on 1 October and 3.17.0 on 29 October, all in the 90 days from 3 August (20). Release notes with a feed, Google Cloud support and public issue trackers for each client library. We did not read the trackers (10 of 15). Official client libraries in eight languages, with Python and Node.js current (15). Python declares 3.10 to 3.15 and Node.js 22 or later. CI results were not checked (5 of 10).

Docling 77

About 25 MCP tools in the default groups (15), selectable as conversion, generation and manipulation groups, with anchor tools to read one part of a document (22 of 25). max_num_pages, max_file_size and page ranges on conversion, Markdown or JSON export (15 of 20). Errors are exceptions with messages, such as an unknown document key listing the valid ones (12 of 20). readOnlyHint and destructiveHint on the MCP tools, and conversions are repeatable (18 of 20). Sensible defaults and one call to convert; Python only, with a docling-client package for docling-serve (10 of 15). 22+15+12+18+10 = 77.

Google Cloud Document AI 70

API reading. fieldMask picks top-level and page fields of the response, imagelessMode removes page images, and page selectors limit what is read. Without a mask the Document carries every token with geometry (20 of 25). List calls page with pageSize and pageToken, operations take a filter, and Layout Parser takes a chunk size. An online request stops at 15 pages, or 30 with imagelessMode, and longer files go through a batch job and Cloud Storage (15 of 20). Errors follow Google's status and message model, with no Document AI error page found (12 of 20). No idempotency key. Failed requests are not billed and processing changes no state, while a repeated successful call bills again. Batch operations can be polled (12 of 20). A processor has to be created before the first call and the host depends on its location. Client libraries in eight languages (11 of 15).

Docling 85

Local-software lines, graded on the library with docling-serve and docling-mcp. Installs from PyPI (docling 2.148.0, per the PyPI release feed) for Python 3.10 and later on macOS, Linux and Windows (20). Public CI workflows for checks, tests and heavy examples; the pass state of the default branch was not read (20 of 25). About 719 open issues and 223 open pull requests on the repository page against 13 releases in four weeks, many of them fixes; replies not sampled (15 of 25). Semantic-release changelog with Feature and Fix sections and Breaking sections where they apply (15). Version 2.x since 2024 (15). 20+20+15+15+15 = 85.

Google Cloud Document AI 90

Hosted reading. Google Cloud status page with a JSON incident feed (20). The feed holds eight incidents back to February 2026, four of them since 3 August, and none names Document AI. They are the 20 August outage that lists 27 products, the 1 September network degradation in us-central1-b, elevated upload and write errors in Cloud Storage in us-central1 from 8 to 9 October and the europe-west2 incident of 20 October (06:00 to 12:08 UTC, 38 products). Batch processing reads and writes Cloud Storage, but the GCS incident was in one region and Document AI isn't listed. The page shows only broad incidents (30). Quotas published with numbers, 1,800 requests a minute per user, 120 online process requests a minute per processor type in us and eu and 5 concurrent batch requests, on a page last updated 28 October (15). No retry or backoff guidance was found on the quotas, limits or request pages. Processing changes no state and failed requests are not billed, so a retry is safe (5 of 15). 99.9 per cent monthly SLA for online and batch prediction, re-read on 1 November (10). v1 is GA (10).