ADR 007: Implicit Previous Output (No Explicit Variable)

ADR 007: Implicit Previous Output (No Explicit Variable)

Status: Accepted Date: 2024-12-27

Context

The original design had users write {{ previous }} in templates to access the prior output of an LLM block:

{% llm id="analysis" %}
Previous: {{ previous }}
New data: {{ ref('sources/data') }}
{% endllm %}

This requires template authors to remember to include {{ previous }} and format it correctly.

Decision

Remove the explicit {{ previous }} variable. Instead, LLM calls with stable IDs automatically receive their prior output as part of the prompt context. The user never writes {{ previous }}.

How It Works

1. LLM call has a stable ID (manual like id='features' or auto-generated)
2. Cache lookup by ID finds previous output
3. If cache hit (same input hash): return cached output, no LLM call
4. If cache miss (input changed): call LLM with previous output in context

# For | extract() filter:
{{ content | extract('feature requests', id='features') }}

# Internally, LLM prompt includes:
# "Previous output for this extraction (maintain if still accurate):
#  {cached_output}"

For {% llm %} Blocks

{% llm id="analysis" %}
Analyze this data: {{ ref('sources/data') }}
{% endllm %}

{# LLM automatically sees its previous output for id="analysis" #}

Rationale

1. Simpler templates: No {{ previous }} boilerplate
2. Consistent behavior: All LLM calls with IDs work the same way
3. Less error-prone: Can't forget to include previous output
4. Separation of concerns: Stability is a caching concern, not a template concern

Consequences

• Template authors don't control how previous output is formatted
• The caching layer is responsible for stability prompts
• Templates are cleaner and more focused on the core logic