What the AI can do

The AI assistant is built on Gemini 3 Flash, with read access to your full data and a set of tools for taking action. This page documents what it can and can't do.

Capabilities by category

Read your data

The assistant can pull from any of these:

• Food log — entries by date range, meal type, with full macros and micros.
• Daily nutrition totals — calories/protein/fat/carbs aggregated per day.
• Weight log — every weigh-in.
• Custom measurements — every metric entry (waist, hip, body fat, etc.).
• Dose log — every dose with compound, value, date.
• Compounds — your enabled compounds, half-lives, kinetics.
• Symptom log — every symptom rating with severity.
• Saved meals — the templates you've created.
• Exercise log — every workout.
• Water log — daily totals.
• Bloodwork — every lab value you've entered.
• Genetics — your variant entries.
• Active conditions — diagnosed conditions you've enabled.
• Menstrual state — current cycle day, phase, predictions.
• Fasting state — current fasting status.
• Day status — tracked / untracked overrides.
• User settings — targets, units, preferences.

For each, you can ask date-ranged questions: "What did I eat last week?", "Show me my weight trend since I started Tirzep."

Run analysis

The assistant has tools that compute live, not just retrieve:

• Correlate any two series — Pearson's r and lag analysis between any two tracked series.
• Partial correlate — same, controlling for one or more series (e.g., "protein × weight controlling for calories").
• Rank correlations — find the strongest correlations against any target series.
• Detect change points — find regime shifts in a single series.
• Compare time windows — before/after analysis for a date range.
• Project a series — extrapolate forward, optionally to a target value.
• Get series daily — raw values for any series, for ad-hoc analysis.

Search the web

For lookups not in your data — drug information, nutrition databases, research. The assistant uses Google Search and cites sources.

Answer support questions

The assistant can pull this help site's full documentation on demand, so you can ask it how Protokol itself works:

• "How do I turn on fasting tracking?"
• "Where do I change my calorie target?"
• "What's the difference between earn and baseline energy mode?"
• "Does Premium include photo logging?"

It fetches the docs live and answers from them rather than guessing, so the response stays in sync with the actual app.

Add or edit entries

Premium and Unlimited tiers can write:

• Log food entries — by name, by photo, or natural-language description.
• Create custom food items — add to your library.
• Search food items — across your library, USDA, OFF.
• Propose food entries — for review and one-tap acceptance.
• Propose bloodwork updates — when you paste lab results.

Read your insights

The dashboard's pattern insights are precomputed and embedded into the assistant's context. So even free-tier accounts (which can't make tool calls) get pattern-level reasoning grounded in real findings.

Examples that work well

Photo-of-plate logging:

(attach photo) "Lunch."

AI: I see chicken thighs, rice, broccoli, and what looks like a slice of cheese. Logging to lunch. ↳ tool: search_food_items(chicken thighs grilled) ↳ tool: search_food_items(white rice cooked) ↳ tool: search_food_items(broccoli steamed) ↳ tool: log_food_entry(chicken thighs, 6 oz, lunch) ↳ tool: log_food_entry(white rice, 1 cup, lunch) ↳ tool: log_food_entry(broccoli, 1 cup, lunch) Logged. Total: 540 kcal, 38g protein, 18g fat, 52g carbs.

Patterns:

"Why does my nausea peak two days after a dose?"

AI: Looking at your dose log and nausea entries... Tirzepatide rises ~36 hours post-injection (peak plasma at hour 24–48). Your nausea data shows the same window — r=0.42 over 47 paired days. The mechanism: GLP-1R activation slows gastric emptying, peaks aligned with peak plasma. This is consistent with the published Tirzepatide PK literature.

Hitting numbers:

"I've eaten 1,400 kcal so far. What should I have for dinner to hit 80g protein?"

AI: You're at 52g protein, target is 130g. So 78g protein remaining and ~400 kcal. Suggestions from your library:

• Greek yogurt + protein powder shake (~30g protein, ~200 kcal).
• Grilled chicken breast 6 oz + steamed vegetables (~50g protein, ~300 kcal).
• Tuna salad on greens (~35g protein, ~250 kcal). Combine the chicken + a protein shake to hit ~80g exactly.

Editing past entries:

"Move yesterday's pasta from lunch to dinner."

AI: Updated. The entry "Spaghetti, 1 cup" moved from lunch to dinner. Yesterday's totals unchanged.

Correlation lookups:

"Is there a relationship between my sleep quality and weight?"

AI: Running correlation... Sleep quality (custom symptom, 0–10) and weight: r = -0.18 over 28 days, partial r = -0.22 controlling for calories. Weak signal. Direction is consistent with poor-sleep → higher weight, but the effect is small enough I'd want more data before treating it as actionable.

Examples that don't work

Medical advice:

"Should I increase my dose?"

AI: I can describe what's in your data — at your current 5mg dose, your weight trend is X, symptom profile is Y. But I can't recommend a dose change. That's a conversation for your prescribing physician.

Speculative scenarios:

"If I switch to Retatrutide, what would my glucose look like?"

AI: I can't simulate switching — I'd be making numbers up. The simulation only runs on what you've actually logged. Once you've started Retatrutide and logged a dose or two, the simulation will model it.

Changing app settings:

"Turn on the rolling 7-day budget."

AI: I can't change settings for you, but I can tell you where it lives — Profile → Tracking → Confirmation mode. The rolling 7-day budget is on by default for tracked accounts. (Pulled from the help docs.)

Tier differences

Capability	Free	Premium	Unlimited
Read your data	✓ (via insights context)	✓	✓
Tool calls (live data + analysis)	—	✓	✓
Web search	—	2 per message	2 per message
Add/edit entries	—	✓	✓
Photo recognition	—	50/day	Unlimited
Daily messages	5	60	Unlimited
Conversation context	Low	Medium	Extreme
Multi-thread chats	—	✓	✓

Free tier still gets useful answers — the assistant has the pre-computed insights and your basic profile in context. It just can't pull live data or take actions.

What the model is

Gemini 3 Flash. We pick this over alternatives because it combines acceptable cost with real tool-use plus Google Search support in a single request — older flash models forced a choice between user-defined tools and web search.

The model knows nothing about you besides what's in the system prompt + your data + this conversation. There's no persistent memory beyond that. Each message starts fresh with the same context-loading flow.

What about hallucinations?

Limitations are real:

• The model can misread an OFF entry's macros if the underlying data is wrong.
• It can invent a serving size when one isn't in the source.
• It can be confidently wrong about pharmacology details — always check claims against published sources for medication-relevant questions.

Mitigations in the system:

• Tool calls return real data from your DB; the model can't fabricate your own logs.
• Web search results are cited.
• Food entries created by AI are visible in your log and reviewable.

Verify anything that affects health decisions.

Related

• Tool calls explained — what those "Searching..." lines mean.
• Privacy & what AI sees — exact data scoping.
• Limits & context length — daily caps and tier differences.