JSON Repair

Repair malformed JSON and publish one result for each input document attempted. Valid JSON documents return successful results with an empty repair report. Successful JSON text uses compact formatting.

Each result contains recordId, document_index, repaired_json, document, repairs, fully_repaired, and error. repaired_json is valid JSON on success and may contain partial text on failure; document is the parsed JSON value on success and null on failure. JSON values may be objects, arrays, strings, numbers, booleans, or null.

Input

Supply exactly one of these fields:

Field	Description
inputText	One JSON document as text.
inputTexts	An ordered array of JSON document strings, with at most 100,000 entries.
inputUrl	A URL returning one JSON document. HTTP and HTTPS destinations must resolve entirely to public addresses.
inputDatasetId	An Apify dataset ID or local dataset name. Every row must contain a sourceText string; rows are read in storage order.

The total size of source content is limited to 8 MiB in UTF-8, across the entire list or dataset, with at most 100,000 documents. Input envelopes and dataset pages are parsed incrementally under a separate serialized-byte cap, enforcing decoded source bytes and document count before constructing the full list. Local input files are size-checked and admitted before SDK initialization, including files without storage metadata. URL downloads cap total HTTP response bytes at 8 MiB before parsing headers or decoding chunks, including chunk extensions, trailers, and redirects; decoded content has a separate 8 MiB cap. Reads use chunks up to 64 KiB with the reference's 30-second socket timeout. Every redirect revalidates DNS and connects to the checked address; discarded redirect bodies are closed without reading, and at most five redirects are followed. Ambiguous input (multiple source fields) and missing input are rejected. API callers may also provide a raw string or list of strings directly.

Each document is limited to nesting depth 128, 500,000 lexical/decoded items (including object keys), and 10,000 repairs. A conservative 384 MiB run workspace budget may refuse repair of an admitted document; that document receives a partial result and workspace_limit error. No salvage is attempted for incomplete strings or decimal/exponent continuations. Numbers that cannot retain their decimal value through serialization return number_precision rather than a changed value. A lone-surrogate source receives invalid_unicode; its original is preserved with JSON escapes and the batch continues.

Output

The default dataset contains one item per attempted document, including already-valid and unrepairable documents. Dataset rows containing lone surrogates and large dataset rows over 5 MiB are stored in the key-value store as RESULT-<8-digit-index> records; their dataset row contains document_index, recordId, fully_repaired, result_key, and result_manifest_key to locate the result. The manifest key is null when the result fits in a single storage record. Result envelopes use key-value store fragments capped at 7 MiB each.

OUTPUT.json is a summary containing resultPrefix, input_documents, charged_documents, published_documents, fully_repaired_documents, unrepaired_documents, and unpaid_documents. Individual results are stored separately. The key-value store also contains billing journals and recovery records.

The engine first parses strict JSON. Its deterministic repair layer removes BOMs, surrounding whitespace and Markdown fences; removes comments outside strings; converts single-quoted strings and Python literals; quotes unquoted object keys; removes trailing commas and trailing content; and tries at most two missing-container completions. It never evaluates code. Repair locations refer to the unwrapped input. Unsupported escapes, missing values, mismatched brackets, non-finite numbers and truncated strings produce fully_repaired: false, partial text, and an error with details.

For example, {name:'Ada',active:True,} produces {"name":"Ada","active":true} with reports for the quoted keys, converted string/literal, and removed trailing comma.

Reliability and billing

• The price is $0.0001 per attempted document. Already-valid and unrepairable documents count as attempts. There is no per-run charge.
• Each document has its own one-event billing request immediately before its repair attempt. Its result checkpoint completes before the next document is charged; publication groups contain at most 500 results.
• The actor saves a source snapshot before charging. Billing journals record accepted work, and each charged result is checkpointed before later charges. Dataset publication tracks durable progress so replay can reconcile writes without charging again or duplicating published results.
• Large snapshots and results stream into bounded fragments. Streamed part envelopes have parts: null; use the completed manifest's partCount and partPrefix to read them in order. Fragment payloads concatenate into the original serialized result or source record.
• Unpaid source documents are preserved under RECOVERY-UNCHARGED-RAW-SUFFIX with a PRESERVATION-REPORT.
• A document-level syntax or complexity failure produces a result and processing continues. Input admission, billing, or storage failures fail the run.

The actor follows the reference's directory layout, Python Docker base, byte limits, billing journal prefixes and charge-first flow. Deliberate differences are one paid document event without a separate run event; per-document result files with an output summary; streaming large envelopes and publishing each bounded group to stay within the memory budget; and rejecting ambiguous inputs or non-HTTP URLs. These changes fit document repair and its bulk inputs. The optional truncated-string salvage layer is omitted to avoid inventing content.

Run via API

curl -H "Authorization: Bearer $APIFY_TOKEN" \
     -H "Content-Type: application/json" \
     -d '{"inputTexts": ["{\"name\":\"Ada\",\"active\":True,}", "{\"count\": 3}"]}' \
     "https://api.apify.com/v2/acts/<username>~json-repair/runs?waitForFinish=60"

Local development

python3 -m pip install -r requirements-dev.txt
mkdir -p storage/key_value_stores/default
# Put {"inputText": "{\"name\":\"Ada\",\"active\":True,}"} in storage/key_value_stores/default/INPUT.json
APIFY_LOCAL_STORAGE_DIR=./storage python3 -m src.main

Run the test suite with:

$python3 -m pytest -q

The SDK startup integration tests require Apify SDK 4.0.2 installed locally under the test interpreter (Python 3.10 or newer). They use temporary filesystem storage with no metadata and prohibit network connections. These three cases are explicitly skipped when that SDK version is unavailable.