Skip to content

PROMPTS.md

Latest commit

 

History

History
77 lines (55 loc) · 3.11 KB

PROMPTS.md

File metadata and controls

77 lines (55 loc) · 3.11 KB
Raw

Prompt Engineering

Model

Gemini 2.5 Flash — chose it because Pawa IT is a Google Cloud Partner. Free tier covers assessment usage, streaming works out of the box.

System Prompt

Lives in backend/app/services/llm_service.py. Two-shot structure:

  1. Persona — "PawaCloud Assistant", GCP specialist built by Pawa IT. Anchoring to a specific identity keeps responses focused vs. generic.

  2. Tone — senior cloud architect mentoring a colleague. Not stiff, not casual. Africa-aware: regional availability, bandwidth constraints, cost sensitivity.

  3. Format — markdown with headings, bullets, code blocks with lang tags. This renders well in the react-markdown frontend without extra parsing.

  4. Honesty — "if you don't know, say so." Prevents hallucination on pricing details and region availability that change frequently.

Parameters

Param Value Why
temperature 0.7 Technical content needs accuracy but not robotic
top_p 0.9 Standard nucleus sampling
top_k 40 Default, works fine
max_output_tokens 2048 Enough for code examples without burning quota

What I tested

  • General questions (travel docs, Kenyan news) — should handle gracefully, not refuse
  • "Deploy FastAPI to Cloud Run" — should give Dockerfile + gcloud commands
  • "What was the impact of the recent floods in Kenya?" — should handle gracefully, not refuse

Document pipeline prompts

Two system instructions, swapped per-call. Picked gemini-2.5-flash over gemini-2.5-pro because the document path is high-volume per-user; Flash quality on summarize/translate is more than sufficient and cost scales linearly with traffic.

Summarize

You are a precise document summarizer. Output is markdown only. Structure: ## Overview, ## Key Points (bullets), ## Action Items (numbered, if any). Preserve named entities, dates, monetary figures verbatim. Do not invent content. If the document is short, the Overview alone may suffice — omit empty sections.

Chose explicit section headings over freeform output because downstream download renderers (docx, pdf) need stable structure. Asking for a specific table-of-contents shape gives consistent results across input domains (legal contracts, news articles, reports).

Translate

You are a professional translator. Translate the document below into {target_language}. Preserve all markdown structure, headings, lists, tables, inline emphasis, and code blocks. Keep proper nouns, brand names, file paths, and URLs unchanged. Do not summarize, paraphrase, or omit content.

The "preserve markdown / proper nouns / URLs" clauses came after early trials where Gemini happily translated Cloud Run to Exécution dans le cloud and broke fenced code blocks. Explicit constraint fixes both.

Source-language hint is included as a non-binding context line, not a constraint, so the LLM can override when our detect_script returns und on mixed-script documents.

max_output_tokens is lifted from the chat default (2048) to 8192 for document calls — translations of multi-page input would otherwise silently truncate.