Skip to main content

Calories prompt design

Problem​

Calorie estimation used to receive only a single event input. After introducing basic_context(e), the model now receives related events too, but a naïve newline-joined payload makes the task ambiguous: the model can no longer tell which event is the target and which lines are only supporting context.

The prompt must help the model:

  • estimate calories for the target event only,
  • use related events only for disambiguation,
  • avoid summing unrelated context events,
  • return a machine-parseable result.

Design goals​

  1. Preserve the target event as the primary source of truth.
  2. Let context help with ambiguous references like “same lunch” or omitted quantities.
  3. Keep the output contract strict: integer or N/A.
  4. Keep token cost modest.
  5. Avoid prompt designs that encourage prose, chain-of-thought, or unstable formatting.

Approaches considered​

Option A: Short prompt with raw joined context​

Example payload:

food: lunch
text: same as yesterday

Pros

  • Minimal code changes
  • Small token footprint

Cons

  • Does not identify the target event
  • Encourages accidental calorie summation across context events
  • Weak guidance for ambiguous references

Option B: Longer system prompt, but still raw joined context​

This improves instructions but keeps the user payload unstructured.

Pros

  • Better than Option A for output discipline
  • Still simple to implement

Cons

  • The model still has to infer which line is the actual target event
  • Important distinctions live only in instructions, not in the data layout

Option C: Structured target/context prompt with strict output contract​

Example payload:

Target event:
food: sandwich

Basic context (related events for disambiguation only):
1. text prep #lunch
2. text same bread as yesterday

Pros

  • Makes the target event explicit
  • Clearly separates context from the event being scored
  • Supports disambiguation without encouraging over-counting
  • Easy to test
  • Low implementation risk

Cons

  • Slightly larger prompt than raw joined text

Option D: Few-shot prompt with many examples​

Pros

  • Could improve edge-case accuracy

Cons

  • Higher token cost
  • Harder to maintain
  • Risks overfitting to examples
  • Not necessary for the current scope

Decision​

Choose Option C.

It gives the model the most important missing signal — which event is the target — while staying compact and deterministic. It also maps naturally onto the new graph dependency basic_context(e) -> calories(e).

Chosen prompt design​

System prompt responsibilities​

The system prompt should:

  • define the target/context distinction,
  • say that calories are for the target event only,
  • allow context to resolve ambiguity,
  • forbid counting separate context events,
  • preserve the integer | N/A response contract.

User payload responsibilities​

The user payload should:

  • show the target event in its own section,
  • list related context events separately,
  • explicitly label context as disambiguation-only.
  • be rendered inside the AI module from the raw target event plus raw context events, rather than pre-rendered upstream.

Expected benefits​

  • Better handling of references like “same lunch”, “another cup”, or preparation notes near the target event.
  • Lower risk that the model totals calories across multiple context events.
  • Better testability because the prompt format is explicit and stable.