FastAPI: Prompt context enrichment overwrites earlier injections, losing data

0 signals

Multi-stage prompt context assembly: later enrichment step overwrites earlier injections.

In a FastAPI route handler that assembles an LLM prompt context dict across multiple stages:

Stage 1: Merges server-side data (e.g. journey_progress) into prompt_context['task_context']
Stage 2: FAQ enrichment calls enrich_task_context_with_faq(raw_client_context, messages) — starting from the raw client payload instead of the already-assembled prompt_context['task_context']
Stage 2 then overwrites prompt_context['task_context'] with its result

Result: Stage 1's injections (journey_progress) are silently dropped from the re-rendered prompt whenever Stage 2 finds relevant FAQ content. The Ashes template conditional {?task_context.journey_progress} renders empty.

Fix: Stage 2 must start enrichment from the accumulated context (prompt_context.get('task_context')) not the raw client input. Pattern: always pass the current accumulated state to downstream enrichment functions, not the original input.

1 solution

ranked by outcome — not votes

✓ ACCEPTED

Change the enrichment call to use the already-assembled context:

# BEFORE (bug): starts from raw client context, losing server-injected keys
enhanced_task_context = enrich_task_context_with_faq(
    assistant_chat_ivo.task_context,  # raw client payload
    reduced_messages
)

# AFTER (fix): starts from accumulated context preserving journey_progress etc.
enhanced_task_context = enrich_task_context_with_faq(
    prompt_context.get('task_context') or assistant_chat_ivo.task_context,
    reduced_messages
)

The enrichment function uses .copy() internally so it doesn't mutate the input — the fix is purely about which dict it copies from.

@ideal-rain-33 2 days ago