> ## Documentation Index > Fetch the complete documentation index at: https://docs.lilfella.app/llms.txt > Use this file to discover all available pages before exploring further. # Troubleshooting > Diagnose installation, workspace, model, query, and Python problems. # Diagnose the failing layer Start with the smallest direct check. A source problem, provider problem, and interpretation problem need different fixes. | Symptom | First check | Likely action | | ------------------------ | ----------------------------- | ---------------------------------------------------------------------- | | A file is missing | `/files` | Check format, hidden path, scan depth, permissions, and `.fellaignore` | | A total looks wrong | `/schema ` | Check type inference, units, casing, nulls, and totals rows | | The model cannot answer | `/login`, then `/model` | Sign in again or choose a tool-capable model | | A query times out | Inspect working or use `/sql` | Add a date filter, narrower columns, grouping, or `LIMIT` | | Python fails | Inspect stderr in working | Simplify the calculation or use SQL | | A PDF has no useful text | Open the PDF outside Fella | OCR it and reopen or `/reindex` the folder | | You typed `/mcp` | `/mcp` | It is an inert experimental command in this release | | Changes do not appear | `/reindex`, then `/files` | Refresh the catalog after external edits | ## A file does not load 1. Run `/files` and read the skipped-source reason. 2. Confirm the extension is supported. Word documents, images, email files, HTML, XML, and OpenDocument files are not current inputs. 3. Confirm the path is not hidden and is within eight levels of the opened folder. 4. Check that `.fellaignore` does not match its name or workspace-relative prefix. Patterns are not globs. 5. Confirm Fella can read the file and that another program is not locking it. 6. For `.parquet` or `.pq`, use a custom DuckDB build or convert the file to CSV/JSON. Release builds do not include Parquet. `fella.md` at the workspace root is intentionally omitted from `/files`; it is prompt context. ## The table loaded incorrectly Run: ```text filename="fella commands" theme={null} /schema table_name ``` Inspect column types and notes. CSV/TSV ingestion can infer headers, skip preambles, coerce common written amounts, normalize named-month dates, and leave out a recognizable trailing total row. These heuristics can be wrong for unusual exports. Clean the source, convert it to a simpler format, or state the exact column and conversion needed. Delimited files stop loading after 2,000,000 rows by default and report truncation. Query results shown to callers are capped at 1,000 rows; use aggregates rather than expecting a full export. ## Fella cannot reach the model Use `/auth` to inspect login state, `/login ` to replace the key, and `/model` to refresh current model IDs. Provider catalogs and IDs can change independently of Fella. HTTP `429` and transient server failures may be retried, but quota, billing, privacy settings, unsupported tool calling, or an unstable endpoint still require provider-side action. ## An answer is plausible but wrong Open the working fold and check, in order: 1. Correct files and tables. 2. Correct date range, units, category spelling, and status filters. 3. Correct join keys and join type. 4. Correct aggregation and denominator. 5. Rows that actually support the prose. 6. Verification warnings. A query can run successfully and return a grounded number while answering the wrong interpretation. Restate the rule explicitly, ask Fella to show intermediate totals, or run direct `/sql`. A green check is not a semantic correctness guarantee. ## There is no working fold Fella can answer general or conversational questions without a tool call. No call means no workspace evidence. Ask a source-bound question such as "Using `transactions.csv`, calculate..." or request the relevant file/table explicitly. ## Python fails or feels unsafe `run_python` is built into Fella. It supplies `median`, `stdev`, Pearson correlation, and simple linear regression helpers, plus `sql(query)`, which returns a list of dictionaries from the mounted workspace. It does not need a Python installation or third-party packages. Python runs without a confirmation prompt inside an embedded RustPython/WASM guest. The guest has no filesystem, network, environment, or subprocess access; the host provides only captured output and bounded read-only SQL. Fuel, memory, stack, source, output, row, and response limits apply. If a calculation can be written clearly in SQL, prefer SQL because its evidence is easier to inspect. ## A PDF cannot be searched Fella extracts an existing text layer and does not perform OCR. If selecting text in a PDF viewer does not produce meaningful text, OCR or convert the document first. Password-protected, malformed, or image-only PDFs may be skipped or produce poor extraction. ## The experimental MCP command `/mcp` is retained as an inert signpost. It does not connect to a service, store connector credentials, or add tools in the personal release. A custom fork that implements a connector owns its network, credential, and permission review. ## The Linux window renders incorrectly On WSLg, if the window opens but layout or compositing is broken, try: ```bash filename="terminal" theme={null} WEBKIT_DISABLE_DMABUF_RENDERER=1 pnpm tauri dev ``` This development workaround does not fix every Linux display failure. Linux requires WebKitGTK 4.1 and the Tauri system libraries listed in the [developer setup](https://github.com/Avijit-Kumar-GIT/fella/blob/main/docs/DEV_SETUP.md). ## Next steps Confirm whether the behavior is outside Fella's current scope. Include the OS, steps, expected result, actual result, and safe sample data.