The short answer
A changed source and a refreshed retrieval index are separate events.
Diagnose stale retrieval by separating original files, derived indexes and cached answers, then verify the repair without deleting useful source history.Follow the old fact through the system
You correct a deadline in a note, then the agent gives you the old date again. Check what it received before blaming the model. The saved file may be current while an index, summary or open conversation still carries the old value.
Check the saved source, the retrieval index and the current conversation separately. A change to one may not have reached the others.
Locate the first stale layer
| Layer | Check | What the result means |
|---|---|---|
| Original file | Open the saved file on the device running retrieval. | Confirm the edit exists in the right copy. |
| Imported material | Find the imported record and its source identity. | Check that the new revision was included. |
| Retrieval index | Run the supported retrieval query separately. | Inspect which passage is returned. |
| Agent context | Start a fresh session using the current result. | Avoid relying on an old answer in chat history. |
Do not rebuild everything before preserving a small record of the failure. Otherwise you may fix the symptom without learning which layer was stale.
Use a distinctive test edit
In a test copy, change a fictional project marker from "version cedar" to "version juniper." Use a phrase unlikely to occur elsewhere. Run the supported import and indexing steps, then query both markers.
This does not prove deletion from backups or providers. It checks whether the active retrieval path reflects a particular source change. Record the source version, update command, query and returned passage so you can repeat the check.
Rebuild derived data through the supported interface
Indexes are implementation-specific. SQLite's FTS5 documentation, for example, distinguishes index maintenance and rebuilding for its supported table configurations. That does not make a raw database command safe for another product.
Use the recovery procedure supplied with your installed release. Keep originals and a recovery copy before applying a repair. Do not delete arbitrary database files because an online article uses the word "reindex." A file may contain authoritative state as well as disposable search data.
Check a repaired answer against an unrelated case
After the marker query is correct, rerun a normal question from the same project and one from another project. A repair that only works for the test phrase may still have coverage or scope problems.
If the index returns the current passage but the answer stays old, focus on the agent's context and reasoning. If the source file itself stays old, inspect synchronization and which device or folder the agent uses. Avoid changing several layers at once.
Turn the incident into a maintenance check
Add the changed-source case to your pilot checklist. Repeat it after changes to source imports, indexing settings or adapters.
Gigabrain separates original knowledge from derived retrieval material. Review the mechanism and supported setup checks before choosing a package. With a local system, you control the files and handle their upkeep.
Trace the recorded schedule correction
Our isolated demonstration first retrieved an inspection date of 18 October. The fixture then changed its disposable source to 25 October.
| Observed step | What it establishes |
|---|---|
| The old receipt failed. | The previous citation no longer matched the source. |
| Maintenance removed stale support. | The earlier graph path could not supply current evidence. |
| Updated support restored the path. | Retrieval reached the revised schedule passage. |
Inspect the actual receipts and maintenance output. The fixture authors its updated relationship support. It does not measure automatic model interpretation of a correction.
Keep originals and add amendments in normal work. Direct mutation of the disposable file is a deliberate failure test.
Record the layer before applying a repair
- Save the current query and returned passage.
- Open the source version on the retrieval device.
- Inspect the import and index status.
- Check the graph support separately.
- Run the documented update for the affected layer.
- Repeat the query.
- Test a new agent conversation separately.
Use the staleness record to retain the failed and repaired states.
Method note: Gigabrain Labs prepared this material with AI assistance. Fictional exercises and executed component checks have separate labels.
