PineScript Expert

A Retrieval-Augmented Generation (RAG) AI Agent built with Pydantic AI that serves as an expert on Pine Script v6, TradingView's programming language for custom indicators and strategies. This agent leverages vector search technology and large language models to provide accurate, context-aware answers and generate working code examples.

Features

• Comprehensive Pine Script Knowledge: Access the entire Pine Script v6 documentation through natural language queries
• Code Generation: Creates custom indicators and strategies based on user requirements
• Interactive Interfaces: Multiple ways to interact with the expert:
  • Web-based UI built with Streamlit
  • Interactive command-line interface
  • Single query execution for scripting
• Multi-Provider Support: Use either OpenAI or OpenRouter models as the LLM backend
• Vector Search: Utilizes pgvector for efficient semantic retrieval of relevant documentation
• Full Documentation Processing: Custom crawler that processes and analyzes TradingView's Pine Script documentation
• Persistent Chat History: Remember conversation context in the Streamlit UI
• Extensive API: Integrate with other systems using the Pydantic AI based architecture

Screenshots

<table> <tr> <td><img src="https://raw.githubusercontent.com/FaustoS88/Pydantic-AI-Pinescript-Expert/main/assets/asset1.png" alt="Asset 1" width="400"></td> <td><img src="https://raw.githubusercontent.com/FaustoS88/Pydantic-AI-Pinescript-Expert/main/assets/asset2.png" alt="Asset 2" width="400"></td> </tr> </table>

Prerequisites

• Python 3.9+
• PostgreSQL with pgvector extension
• OpenAI API key (required for embeddings and default LLM)
• OpenRouter API key (optional, for alternative LLM providers)
• Docker (optional, for running PostgreSQL with pgvector)

Quick Start

1. Clone the repository

   git clone https://github.com/FaustoS88/Pydantic-AI-Pinescript-Expert.git
   cd pinescript-expert
   

2. Setup PostgreSQL with pgvector using Docker

# Create a directory for Docker volume if it doesn't exist
mkdir -p ~/pinescript_postgres_data

# Run PostgreSQL with pgvector on port 54322 (different from standard 5432)
docker run --name pinescript-pgvector \
  -e POSTGRES_PASSWORD=postgres \
  -p 54322:5432 \
  -v ~/pinescript_postgres_data:/var/lib/postgresql/data \
  -d pgvector/pgvector:pg16

3. Install requirements and set up the environment

   python setup.py
   # Edit the created .env file with your API keys
   

4. Initialize the database

   python init_db.py
   

5. Populate the database with Pine Script documentation

   python pinescript_crawler.py
   

6. Start using the Streamlit UI

   streamlit run streamlit_ui.py
   

   Or, for CLI interface:

   python interactive.py
   

Usage Examples

Web Interface

Start the Streamlit interface to interact with the agent through a web UI:

streamlit run streamlit_ui.py

Command Line Interface

Launch an interactive shell for conversational access to the agent:

python interactive.py

Example session:

=================================================================
 Pine Script Expert Agent - Interactive Shell 
=================================================================
Ask any question about Pine Script v6 or type 'exit' to quit.
Type 'clear' to clear the conversation history.
=================================================================

> How do I create a simple moving average crossover strategy?

Processing your question...

================================================================================
To create a simple moving average crossover strategy in Pine Script v6, you'll need to:

1. Set up your indicator or strategy
2. Calculate two moving averages of different lengths
3. Determine crossover conditions
4. Generate buy/sell signals
5. Optionally add plotting for visualization

Here's a complete example:

```pine
//@version=6
strategy("Simple MA Crossover Strategy", overlay=true)

// Input parameters
fastLength = input.int(9, "Fast MA Length")
slowLength = input.int(21, "Slow MA Length")

// Calculate moving averages
fastMA = ta.sma(close, fastLength)
slowMA = ta.sma(close, slowLength)

// Determine crossover conditions
buySignal = ta.crossover(fastMA, slowMA)
sellSignal = ta.crossunder(fastMA, slowMA)

// Execute strategy
if (buySignal)
    strategy.entry("Buy", strategy.long)
    
if (sellSignal)
    strategy.entry("Sell", strategy.short)

// Plot moving averages
plot(fastMA, "Fast MA", color=color.blue)
plot(slowMA, "Slow MA", color=color.red)

// Plot buy/sell signals
plotshape(buySignal, "Buy Signal", shape.triangleup, location.belowbar, color.green, size=size.small)
plotshape(sellSignal, "Sell Signal", shape.triangledown, location.abovebar, color.red, size=size.small)

Key components explained:

• We use ta.sma() to calculate the simple moving averages
• ta.crossover() and ta.crossunder() detect when the fast MA crosses above or below the slow MA
• strategy.entry() executes buy and sell orders when crossovers occur
• plot() and plotshape() visualize the MAs and signals on the chart

You can customize this by changing:

• MA types (SMA, EMA, WMA, etc.)
• Length parameters
• Adding additional conditions
• Adding stop-loss and take-profit levels ================================================================================

### Single Query Mode

Use the agent for a one-time query:

```bash
python run.py query "How do I calculate RSI in Pine Script?"

Database Inspection

The project includes tools to inspect the vector database:

# Count documents in the database
python db_inspect.py count

# List document titles (first 20)
python db_inspect.py list

# View a specific document
python db_inspect.py view 508

# Test search functionality
python db_inspect.py search "how to use request.security for different timeframes"

RAG Pipeline — Evaluation Results

The retrieval pipeline has been systematically improved across two tiers and measured with RAGAS (40-question test set, categories: function lookup, conceptual, code generation, complex multi-concept).

| Pipeline | Faithfulness | Context Relevance | vs Baseline | |----------|-------------|-------------------|-------------| | Baseline (flat chunks, L2 search) | 0.779 | n/a | — | | Tier 1 (hybrid search, MMR, recursive chunking) | 0.774 | — | -0.6% | | Tier 2 (+ Anthropic Contextual Retrieval) | 0.833 | 0.919 | +6.9% |

Code generation improved from 0.51 → 0.64 (+21%) with Tier 2. Context Relevance 0.919 means the retriever finds the right chunks 92% of the time — the remaining gap is a content problem (docs lack complete strategy templates), not a retrieval problem.

Tier 1 improvements (docs): hybrid BM25+vector search, similarity threshold, cross-encoder reranking, recursive chunking with overlap, MMR deduplication, contextual chunk headers

Tier 2 improvements (docs): code-aware chunking (fenced blocks never split), Anthropic Contextual Retrieval (LLM prefix per chunk at crawl time), content type detection, metadata columns

Evaluation details: Tier 1 | Tier 2

Running RAGAS Evaluation

# Tier 1 retrieval (hybrid + MMR)
python tests/ragas_eval.py --retrieval tier1 --output results/tier1_YYYYMMDD.json

# Baseline retrieval (L2 only)
python tests/ragas_eval.py --retrieval baseline --output results/baseline_YYYYMMDD.json

Running the Contextual Re-Crawl (Tier 2 activation)

# Standard re-crawl (code-aware split, free)
python pinescript_recrawl_light.py --clear

# Contextual re-crawl (LLM prefix per chunk — ~$8, ~2h for 4,910 chunks)
python pinescript_recrawl_light.py --contextual --clear

---

Key Components

• agent.py: Core agent implementation with RAG capabilities
• pinescript_crawler.py: Documentation crawler and vector database population
• db_schema.py: Database schema definitions
• streamlit_ui.py: Web-based user interface with persistent chat history
• interactive.py: Command-line interface
• run.py: Convenience runner for various operation modes
• init_db.py: Database initialization
• clear_database.py: Database cleaning utility
• db_inspect.py: Database inspection tools

Advanced Configuration

Model Presets

Switch between LLM providers using the --model flag. Presets are defined in config.py:

| Preset | Model | Temperature | Max Tokens | |-----------|---------------------------------|-------------|------------| | default | openai/gpt-4.1-mini | 0.2 | 2000 | | codex | openai/gpt-5.3-codex | 0.1 | 4096 | | opus | anthropic/claude-opus-4-6 | 0.2 | 4096 | | flash | google/gemini-3-flash-preview | 0.3 | 2000 |

# Use a preset
python run.py query "How to use request.security?" --model codex
python run.py interactive --model opus

# Or pass any OpenRouter model ID directly
python run.py query "Explain ta.sma()" --model "anthropic/claude-sonnet