09 / Splitline field manual
Troubleshoot Splitline
Use the notice first: Splitline distinguishes invalid text, parse evidence, working-limit stops, read-only safety, visualization diagnostics, save conflicts, and model failures so the remedy can stay narrow.
Updated
Start with the symptom
| Symptom | First action |
|---|---|
| “The file is not valid UTF-8 text” | Re-export or convert the source as UTF-8 without replacing undecodable bytes. |
| Valid rows appear beside issue counts | Open Issues; Splitline recovered the valid records and kept the bad source locations visible. |
| “Opened part of…” and a read-only badge | Read the truncation reason. Raise the Finder byte budget only when size caused the stop, then reopen. |
| Raw or Raw JSON is unavailable | Use the bounded workspace, or verify the view-specific limits below. |
| Save says the file changed on disk | Do not force an overwrite. Preserve the Splitline edits, reopen the current disk version, and reconcile. |
| Visualize keeps the prior chart and names a field or limit | Keep the named selection visible, choose a compatible or lower-cardinality field, or narrow the current Rows scope. |
| Analyze rejects or cannot reach the endpoint | Check the URL policy, authentication, model ID, server state, and certificate setting. |
Fix UTF-8 errors
Splitline decodes all supported formats strictly as UTF-8. It does not substitute the replacement character for undecodable bytes because that would silently change a field or record.
Re-export the file from its producing tool with UTF-8 encoding, or convert it with a tool that reports invalid input. A leading UTF-8 byte order mark is supported. Changing a filename extension does not change its encoding.
A Finder row file cut by the byte budget is trimmed back to a safe newline and UTF-8 boundary before decoding. If Splitline still reports invalid UTF-8, the invalid bytes are inside the retained file content rather than an artifact of the size cut.
Read parsing issues
Parsing behavior depends on the format:
- JSON documents must parse as one complete JSON value. Invalid syntax prevents the document from opening.
- JSONL expects one top-level object on each non-empty line. Malformed JSON, non-object values, unsafe integers, and numeric overflow or underflow become per-line issues; other valid objects remain available.
- CSV and TSV use the first non-blank row as the header. A row with the wrong number of columns or a file ending inside an unterminated quoted value becomes an issue rather than being guessed into shape.
Delimited typing is conservative. Quoted fields always stay text. Unquoted 007, 0.10, 1e5, phone-like strings, and unsafe integers also stay text; a token becomes a number only when converting it back reproduces the source exactly. A bare empty field is null, while an empty quoted field is an empty string.
Open Issues to see source locations and schema checks. Correct the producing data or, for a complete row file, edit valid records in Splitline. Malformed source that never became a record remains untouched by unrelated saves.
Resolve partial loads
Finder-opened row datasets stop cleanly at the first working budget they reach:
- the adjustable byte limit, 32 MiB by default and 1–256 MiB in Settings;
- 250,000 records;
- 100,000 parse issues; or
- 2,000,000 top-level field values.
The notice says how much loaded and why. JSONL byte prefixes end on a complete record. CSV and TSV prefixes end on a newline and valid UTF-8 boundary; because quoted values can span lines, the final prefix may still report one unterminated-quote issue.
If the byte budget caused the stop, choose the top-bar gear, raise Finder dataset size limit, and reopen the file. The change applies only to the next open and cannot exceed the native 256 MiB file cap. If a record, issue, or field-value budget caused the stop, raising the byte setting will not help; inspect the representative prefix or reduce the input outside Splitline.
Finder JSON behaves differently: it must load whole and is refused rather than truncated beyond 16 MiB, 100,000 values, or 256 levels. Browser-selected files are read fully in browser memory.
Explain read-only state
A row file is read-only when Splitline did not retain complete source text. This prevents a loaded prefix from being saved over the unseen tail of the real file. Rows, Schema, Issues, search, and inspection remain available; editing, deletion, Raw, and Save do not.
A complete row file larger than 1 MiB can still be editable even though Raw is unavailable—the Raw surface has its own tighter rendering limit. Likewise, a JSON document may fit Raw JSON but remains view-only by product design.
Raw needs a complete retained row source at or below 1 MiB. Raw JSON needs a complete parsed document at or below 1 MiB, 5,000 values, and 64 levels. Neither threshold grants permission to write a JSON document.
Resolve save problems
For a Finder-opened row file, Splitline records the size and modification time at load. If either changed before Save, it refuses the write with a conflict message. The Splitline edits stay in memory.
First stop editing the file in other apps. Copy or record the changed rows you need from the Splitline session, then remove and reopen the current disk file and reapply the intended changes. This explicit reconciliation avoids erasing either version. If the file did not intentionally change, check whether a sync tool or formatter is rewriting it.
For a permission or storage error, restore write access or available disk space and retry while the session remains open. Cancelling Export copy is not an error and leaves every unsaved change available for another attempt.
Fix Visualize
Visualize preserves the last valid chart when a manual setting becomes incompatible. The selected field stays visible so the repair is explicit: a category field is needed for grouping, numeric fields are needed for scatter axes and numeric aggregation, and a metric needs a numeric measure. Choose a compatible field or restore the local/agent source plan; Splitline never silently swaps in another column.
A cardinality diagnostic means the local scan reached a recoverable working bound: 50,000 distinct series values, 50,000 distinct group values, or 100,000 group-by-series cells. Choose a lower-cardinality field or use Use current Rows scope after narrowing Rows. Arithmetic diagnostics mean finite inputs produced an unrepresentable sum or range; change the measure or aggregation rather than trusting a clipped result.
If a chart chunk fails to render, use its local Retry chart action. The rest of the file workspace and the underlying locally computed result stay available. If evidence changed after an edit, deletion, undo, or revert, reselect the mark: Splitline deliberately clears stale pins and Rows links when row identity may have changed.
Fix Analyze
Check these in order:
- Open Analyze Settings and verify the complete chat-completions endpoint, authentication mode, optional model ID, and saved-connection source.
- Use HTTPS for a remote host. Plain HTTP is accepted only on
localhost,127.0.0.1, or[::1]. - Confirm the model server is running and that the token or custom header has permission to use the chosen model.
- For a trusted self-hosted server with an unverified certificate, edit the connection and decide whether to enable the explicit certificate-risk checkbox. Never enable it merely to silence an unexplained remote TLS error.
- If
SPLITLINE_AI_*variables provision the connection, launch Splitline from an environment that actually contains them. Apps opened from Finder do not inherit shell variables.
If the request reaches the model but the transcript reports an unsupported reply, the provider returned something outside Splitline’s two allow-listed shapes. Rephrase the task, verify that the endpoint supports JSON chat-completions output, or choose another model. No rejected reply is executed.
Contact support safely
For product help, email support@se-ai.live and describe the visible symptom, app version, file format, and the exact non-sensitive error text. Do not attach a dataset or include a file path, field name, field value, prompt, model response, token, endpoint secret, recovery code, checkout claim token, device credential, or signed lease.
For purchase or activation errors, follow the separate License error guide and use Splitline’s Copy diagnostic action. For policies and expected response time, see Support.