BLITZ Lite User Guide

1. Install and launch

1. Install Python 3.10 or later for Windows and include Tcl/Tk.
2. Install Ollama or LM Studio separately. Download a model in that application and start its local API service.
3. Unzip the BLITZ Lite package to a folder you can write to.
4. Double-click START_BLITZ_LITE.bat. If Windows does not associate py, launch lite_console.py with Python 3.10 or later.
5. Choose the Connection tab, choose the provider, check the endpoint, then click Find models and Save connection.

2. Ask and compare

1. In Chat and Compare, select a model. Use Ctrl-click to select up to three primary models.
2. Enter your question or research task and click Send / Run.
3. Each model gets the same prompt. Ordinary chat sends that model's saved conversation turns with it.
4. Review answers and the input/output token and timing details shown below each answer.
5. Start a clean context with New conversation.

BLITZ Lite makes non-streaming requests. The full response appears after the provider finishes. The request cannot be cancelled in this release; stop it in Ollama or LM Studio if the provider exposes that control.

3. Run a small research trial

1. Enter a short experiment name in Research experiment name.
2. Select the models to compare. Set repeats from one to five.
3. Enter one prompt and run it. Each replicate sends the same prompt to the selected model without earlier conversation turns.
4. Open History and Exports to inspect each response and its metadata.
5. Export selected events as CSV or all events as JSONL for later analysis.

Record the model/provider version and other conditions in the experiment name or a separate research note. BLITZ Lite does not claim statistical significance or call a small trial a benchmark.

4. Use the verifier and reviewer

Verifier

When enabled, a verifier request begins as each primary answer finishes. It extracts claims and compares them with the evidence text you supplied. With no evidence, it can only critique plausibility or internal consistency; it cannot independently establish external facts. The result is saved as its own event and labeled by model identity.

Reviewer

The optional reviewer checks for omissions, contradictions, unsupported leaps, counterexamples, overconfident phrasing, and limits. It does not silently alter the primary answer.

Choose different model names when possible. If you use the same model for the answer and review, the record says it is the same model. A different model can still share training data and limitations.

5. Understand the metadata

Token accounting differs by runtime and may include provider-side formatting. BLITZ Lite does not estimate tokenizer counts or show cumulative session tokens as context-window use.

6. History, storage, and exports

Settings and captures live in %LOCALAPPDATA%\NuAegis\BLITZ Lite. Captures are daily JSONL files under captures\YYYY-MM-DD\events.jsonl. Each event includes a schema version, IDs, timestamp, endpoint, model, prompt, answer, evidence indicator/text, status, and provider metadata. JSONL is append-only during normal use.

• Markdown: export the latest event for reading.
• CSV: export selected history rows for a spreadsheet.
• JSONL: export all saved events with full metadata.
• Open data folder: inspect, back up, or manually remove the local data.

Prompts, model responses, and supplied evidence are retained locally because conversation capture is a core feature. Review exports before sharing them.

7. Privacy and safety

• No telemetry, analytics, account, sync, auto-update, MCP, shell, arbitrary code execution, or embedded web access.
• Ollama/LM Studio run separately. BLITZ Lite sends requests only to the endpoint shown in Connection.
• For non-local endpoints the app asks before it sends conversation content. Provider terms govern any remote endpoint.
• Endpoint URLs containing credentials, query parameters, or fragments are rejected so secrets are not written into local records.
• Never paste secrets, passwords, or private data unless you understand the endpoint and storage consequences.
• Evidence is sent to the primary model only when part of the user prompt, and to enabled reviewer/verifier calls.

8. Troubleshooting

• No models found: start the provider service, verify the endpoint and port, and confirm at least one model is installed and loaded/available.
• Connection refused: check the provider's local API toggle and endpoint URL. Then use Find models again.
• Model error: confirm the model identifier exactly matches the provider list and that the machine has enough available memory.
• Token values blank: the provider response did not include usage counts. The blank is intentional.
• Request takes a long time: the provider is still generating/loading or the endpoint is unresponsive. This release has no cancel control.
• History missing: use Open data folder and check the daily JSONL file. Do not edit a file while a run is actively writing to it.

9. Scope and license

This is a compact local-model console, not a web research agent, dataset extractor, tool runner, or fact-certification system. It does not bundle Python, Ollama, LM Studio, or model weights.

BLITZ Lite is released under the MIT License. The full license text is in LICENSE. There is no paid checkout or separate commercial license required by the included license.

Support: remi@nuaegis.com.