Formatting System Prompts

# Formatting System Prompts

This directory contains all **parameter formatting prompt versions** used in the LLM Query Engine.  

Each prompt is stored as a separate JSON file and loaded automatically by `prompt_loader.py`.

## Comparison Table

| Version | Style | Strength | Avg F1 | JSON OK | Perfect matches | Best For |
|---|---|---|---:|---:|---:|---|
| **v1 – default** | verbose baseline | broad coverage, full rules | **0.863** | 96.2% | 53.8% | Early baseline, documentation reference |
| **v2 – strict** | short strict | high precision & JSON correctness | **0.896** | **100.0%** | **61.5%** | Strict evaluation, clean operator mapping |
| **v3 – normalization** | normalization-heavy | strong numeric & fuzzy normalization | **0.896** | **100.0%** | **61.5%** | Handling fuzzy wording + numeric ranges |
| **v4 – hybrid (default)** | v2 + v3 | production-like balance | **0.896** | **100.0%** | **61.5%** | Recommended default for the pipeline |

> Based on 26 test queries (see notebook results). v2, v3, and v4 all achieve the highest F1 and JSON correctness. v4 behaves most like a production default: it combines strict operator handling with robust numeric & fuzzy normalization, so we currently consider **v4 the best candidate for the pipeline default**.

## formatting_v1_default

**Purpose:**  
The default formatting prompt translates extracted raw values into a strict JSON array format that downstream systems can consume.  
It contains comprehensive rule explanations for each parameter type.

**Characteristics:**
- Full rule set for converting raw values into structured JSON
- Dynamic `{allowed_parameters_list}` injection via `prompt_loader.py`
- Handles all schema parameters: rooms, price_euro, surface_m2, listing_type, property_type, district, neighborhood, postal_code, near_* flags, balcony, first_tenant, etc.
- Supports multi-condition outputs (e.g., between X and Y)
- Enforces JSON-only responses with strict formatting rules

**Recommended Use Cases:**
- Baseline formatting behaviour for comparison against stricter or few-shot-driven prompts  
- Ensuring deterministic and well‑structured JSON for model evaluations  
- Early-stage pipeline validation and debugging

**Few-Shot Group:** `basic_v1`

**File:** `formatting_v1_default.json`

**User template:**