PRANAG-AI — Prompt Parser

Converts a natural language prompt into a structured spec.json for the simulation pipeline. Accepts plain English text from a file, CLI argument, or interactive input, and runs it through a 5-node LangGraph pipeline that parses, searches, researches, builds, and validates the output.

---

Project Structure

Prompt-Parser-AI-main/
├── input_folder/
│   └── input.txt              # ← put your prompt here
├── orchestrator/
│   ├── prompt_parser.py       # LLM + regex parsing (Node 1)
│   ├── research_fetcher.py    # Semantic Scholar API (Node 3)
│   ├── spec_builder.py        # assembles final spec dict (Node 4)
│   ├── output_validator.py    # Pydantic validation (Node 5)
│   ├── output_exporter.py     # saves spec.json to disk (Node 5)
│   └── workflow.py            # LangGraph pipeline definition
├── search_engine/
│   ├── vector_store.py        # ChromaDB wrapper + seed traits
│   ├── similarity_search.py   # cosine similarity search
│   ├── embeddings.py          # sentence-transformers model
│   └── data_cleaner.py        # text normalisation utilities
├── shared/
│   ├── config.py              # all settings via .env
│   └── models.py              # Pydantic models (ParsedPrompt, Spec)
├── outputs/                   # generated spec.json files land here
├── chroma_db/                 # ChromaDB persisted on disk
├── main.py                    # CLI entry point
├── app.py                     # Streamlit UI entry point
├── requirements.txt
└── .env.example

---

Prerequisites

• Python 3.11 or 3.12
• Ollama installed and running locally
• DeepSeek-R1:7b model pulled in Ollama

---

Setup

1. Clone and create a virtual environment

git clone <repo-url>
cd Prompt-Parser-AI-main
python -m venv venv

# Windows
venv\Scripts\activate

# macOS / Linux
source venv/bin/activate

2. Install dependencies

pip install -r requirements.txt

3. Copy and fill in the environment file

cp .env.example .env

Open .env and fill in your values:

OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=deepseek-r1:7b
OLLAMA_TIMEOUT=120

# Optional — get a free key at https://www.semanticscholar.org/product/api
SEMANTIC_SCHOLAR_API_KEY=

OUTPUT_DIR=./outputs
EXPORT_TO_FILE=true

# Leave blank until simulation team shares their URL
SRIKAR_ENDPOINT=

4. Pull the LLM model

ollama pull deepseek-r1:7b
ollama serve

---

Running the Pipeline

Option 1 — File input (primary method)

Place your English prompt inside input_folder/input.txt:

wheat for Jodhpur at 48°C with low rainfall, drought resistant

Then run:

python main.py

Windows note: Do not use PowerShell echo to write the file — it saves as UTF-16 which causes a decode error. Instead open the file in Notepad and save it, or use:

[System.IO.File]::WriteAllText("input_folder\input.txt", "your prompt here", [System.Text.Encoding]::UTF8)

Option 2 — CLI argument

python main.py "wheat for Jodhpur at 48°C with low rainfall"

Option 3 — Streamlit UI

streamlit run app.py

Open http://localhost:8501 in your browser, type a prompt, and click Generate Spec.

Input priority order: CLI argument → input_folder/input.txt → interactive terminal prompt.

---

How the Pipeline Works

input_folder/input.txt
        │
        ▼
  [1. parse_node]  ── prompt_parser.py
        │              Regex fallback + Ollama LLM (DeepSeek-R1)
        │              Output: parsed dict {crop, location, temperature, ...}
        ▼
  [2. search_node] ── similarity_search.py + ChromaDB
        │              Embeds query → cosine search over trait database
        │              Output: top-10 relevant trait strings
        ▼
  [3. research_node] ── research_fetcher.py
        │               Queries Semantic Scholar API for recent papers
        │               Output: list of {title, key_finding, url, year}
        ▼
  [4. build_node]  ── spec_builder.py
        │              Merges parsed + traits + research into spec dict
        │              Computes confidence score
        ▼
  [5. validate_node] ── output_validator.py + output_exporter.py
                        Pydantic schema check → saves outputs/spec_*.json
                        Optionally POSTs to simulation endpoint

If validation fails, the pipeline retries from Node 1 up to 3 times. The pipeline never fully crashes — regex fallback guarantees a parsed result even when Ollama is offline, and hardcoded research fallbacks handle Semantic Scholar downtime.

---

Output — spec.json

Every successful run writes a file to ./outputs/ named spec_YYYYMMDD_HHMMSS_<uuid>.json.

Example output:

{
  "crop": "wheat",
  "location": "Jodhpur, Rajasthan",
  "temperature": 48.0,
  "humidity": null,
  "rainfall": 300.0,
  "soil_type": "sandy loam",
  "stress_conditions": ["extreme heat stress", "drought"],
  "target_traits": ["heat tolerance", "drought resistance", "deep root system"],
  "retrieved_traits": [
    "heat shock protein HSP70 expression increases wheat survival above 45°C",
    "deep root architecture in wheat reaching 120cm for subsoil moisture access"
  ],
  "scientific_basis": ["Wheat yield under heat stress reduced by 6% per °C above 30°C..."],
  "research_titles": ["Heat stress tolerance mechanisms in wheat"],
  "research_sources": ["https://api.semanticscholar.org/..."],
  "research_years": [2023],
  "constraints": {},
  "confidence": 0.87,
  "pipeline_version": "1.0.0",
  "generated_at": "2026-05-28T14:45:16.123456+00:00"
}

---

Configuration Reference

All settings are in shared/config.py and can be overridden via .env:

Setting	Default	Description
OLLAMA_BASE_URL	http://localhost:11434	Ollama server URL
OLLAMA_MODEL	deepseek-r1:7b	LLM model name
OLLAMA_TIMEOUT	120	Seconds before LLM call times out
SEMANTIC_SCHOLAR_API_KEY	(empty)	Optional — avoids rate limiting
OUTPUT_DIR	./outputs	Where spec.json files are saved
EXPORT_TO_FILE	true	Set false to skip disk write
SRIKAR_ENDPOINT	(empty)	Simulation team's POST endpoint
CHROMA_PERSIST_DIR	./chroma_db	ChromaDB storage path
SEARCH_TOP_K	10	Number of traits to retrieve
MAX_RETRIES	3	Pipeline retry limit on validation failure

---

Connecting to the Simulation Team

Once the simulation team (Srikar/Aryan) shares their endpoint URL, set it in .env:

SRIKAR_ENDPOINT=http://srikar-sim:8000/spec

The pipeline will automatically POST the validated spec after each successful run alongside saving it to disk.

---

Scaling the Search Engine

The vector store currently has 47 hardcoded seed traits. To load a larger dataset:

from search_engine.vector_store import get_collection
from search_engine.embeddings import embed_batch
import csv, uuid

col = get_collection()
texts, ids, metas = [], [], []

with open("your_data.csv") as f:
    for row in csv.DictReader(f):
        texts.append(row["text"])
        ids.append(str(uuid.uuid4()))
        metas.append({"crop": row["crop"], "domain": row["domain"]})

vectors = [v.tolist() for v in embed_batch(texts)]
col.upsert(ids=ids, documents=texts, embeddings=vectors, metadatas=metas)

The orchestrator's search_node picks up new entries automatically on the next run — no pipeline changes needed.

---

Team

Ashirwad, Om and Shambhavi