RAG Implementation Guide

Overview

This skill provides Retrieval-Augmented Generation (RAG) capabilities for semantic search and knowledge retrieval. It supports vector-based similarity search using FAISS and transformer embeddings, compatible with the agentic4mle project's RAG implementation.

Quick Start

In this project, Operator skill scripts are executed through the use_skill tool's run_script action (see evomaster/agent/tools/skill.py). Therefore, this guide focuses on script invocation as the primary usage pattern.

1) Semantic Search (Recommended Entry: scripts/search.py)

Running Scripts via use_skill

Arguments are passed through script_args as space-separated strings (internally executed as python /abs/path/to/script.py {script_args}).

Example (retrieve top 5 results with optional knowledge field output):

use_skill(
  skill_name="rag",
  action="run_script",
  script_name="search.py",
  script_args="--vec_dir /path/to/vectorstore --query \"What is the main topic?\" --top_k 5 --threshold 1.5 --nodes_data /path/to/nodes_data.json --content_path content.text --output json"
)

Direct Command Line Execution (Equivalent)

python evomaster/skills/rag/scripts/search.py \
  --vec_dir /path/to/vectorstore \
  --query "What is the main topic?" \
  --top_k 5 \
  --threshold 1.5 \
  --nodes_data /path/to/nodes_data.json

2) Encoding Only (scripts/encode.py)

use_skill(
  skill_name="rag",
  action="run_script",
  script_name="encode.py",
  script_args="--text \"What is the main topic?\" --model evomaster/skills/rag/local_models/all-mpnet-base-v2"
)

3) Task Knowledge (basic tools + search.py)

To get data_knowledge and model_knowledge for a task:

1. •Get the query text: use execute_bash to cat <simple_instructions_dir>/<task_name>/simple_instructions.txt from project root, or use str_replace_editor with command view and the absolute path to that file.
2. •Run search.py with --query "<query from step 1>", --top_k 1, --threshold 1.5, --nodes_data pointing to the draft DB (e.g. draft_407_75_db.json), and --output json.
3. •From the JSON result, take the first item in results; its content (or the node in nodes_data for that node_id) contains data_knowledge and model_knowledge for draft-stage DB.

Core Components

1. Vector Database Structure

The vector database follows this structure:

MLE_DATABASE/
├── MLE75_tasksummary_v2.json          # Task summaries JSON file
├── simple_instructions/                # Simple instructions directory
│   ├── <task_name>/
│   │   └── simple_instructions.txt
│   └── ...
└── node_vectorstore/                  # Vector stores (multiple dimensions)
    ├── draft/                          # Draft stage vectorstore (768 dims)
    │   ├── faiss.index                 # FAISS index file
    │   ├── embeddings.npy              # Pre-computed embeddings
    │   ├── nodes.jsonl                 # Node ID mappings (one JSON per line)
    │   └── draft_407_75_db.json        # Full node data with knowledge content
    ├── improve/                        # Improve stage vectorstore
    │   ├── faiss.index
    │   ├── embeddings.npy
    │   ├── nodes.jsonl
    │   └── nodes_for_improve.json
    └── debug/                          # Debug stage vectorstore
        ├── faiss.index
        ├── embeddings.npy
        ├── nodes.jsonl
        └── nodes_for_debug.json

Note: Multiple vectorstore directories exist for different embedding dimensions. Each directory corresponds to a specific model:

• •node_vectorstore/ (768 dims) - Uses local model: evomaster/skills/rag/local_models/all-mpnet-base-v2
• •node_vectorstore_512/ (512 dims) - Uses text-embedding-3-large (512 dimensions)
• •node_vectorstore_768/ (768 dims) - Uses text-embedding-3-large (768 dimensions)
• •node_vectorstore_1024/ (1024 dims) - Uses text-embedding-3-large (1024 dimensions)
• •node_vectorstore_2048/ (2048 dims) - Uses text-embedding-3-large (2048 dimensions)
• •node_vectorstore_3072/ (3072 dims) - Uses text-embedding-3-large (3072 dimensions, default)

Important: Use the vectorstore directory that matches your model's output dimension. The default node_vectorstore/ uses the local model, while other directories use OpenAI's text-embedding-3-large model with different dimensions.

2. Embedding Models

Default model location: evomaster/skills/rag/local_models/all-mpnet-base-v2

Supported embedding models:

• •Local model (default, for node_vectorstore/): evomaster/skills/rag/local_models/all-mpnet-base-v2 - 768 dimensions
• •text-embedding-3-large (for node_vectorstore_* directories): OpenAI embedding model with configurable dimensions (512, 768, 1024, 2048, 3072)
• •sentence-transformers/all-mpnet-base-v2: HuggingFace model, 768 dimensions
• •sentence-transformers/all-MiniLM-L6-v2: Faster, smaller (384 dimensions)
• •Any HuggingFace transformer model compatible with AutoModel

Important:

• •When using a local model path, ensure the model directory contains config.json, model.safetensors, and tokenizer files.
• •When using text-embedding-3-large, ensure you use the correct vectorstore directory matching the dimension used during indexing.

3. Retrieval Methods

Similarity Search

# Search with distance threshold
results = searcher.search_similar(
    query_emb,
    top_k=10,
    distance_threshold=1.5  # Filter results beyond this distance
)

Text-based Search

# Direct text search (encodes and searches in one step)
results = searcher.search_by_text(
    query_text="What is the main topic?",
    top_k=5,
    distance_threshold=None
)

Usage Patterns

Pattern 1: Basic Knowledge Retrieval

use_skill(
  skill_name="rag",
  action="run_script",
  script_name="search.py",
  script_args="--vec_dir evomaster/skills/rag/MLE_DATABASE/node_vectorstore/improve --nodes_data evomaster/skills/rag/MLE_DATABASE/node_vectorstore/improve/nodes_for_improve.json --query \"How to improve model performance?\" --top_k 5"
)

Pattern 2: Multi-Stage Retrieval

# improve
use_skill(skill_name="rag", action="run_script", script_name="search.py",
          script_args="--vec_dir evomaster/skills/rag/MLE_DATABASE/node_vectorstore/improve --nodes_data evomaster/skills/rag/MLE_DATABASE/node_vectorstore/improve/nodes_for_improve.json --query \"...\"")

# debug
use_skill(skill_name="rag", action="run_script", script_name="search.py",
          script_args="--vec_dir evomaster/skills/rag/MLE_DATABASE/node_vectorstore/debug --nodes_data evomaster/skills/rag/MLE_DATABASE/node_vectorstore/debug/nodes_for_debug.json --query \"...\"")

Pattern 3: Custom Knowledge Extraction

Currently, scripts/search.py defaults to reading nodes_data[<node_id>].content.improve_knowledge (consistent with agentic4mle/utils/vectorstore.py).

If you need to read different fields during migration, the recommended approach is:

• •Copy scripts/search.py to a new script (e.g., search_debug.py)
• •Modify get_knowledge() to read fields like bug_fix_specific / bug_fix_abstract / full_plan / code
• •Call it via use_skill(..., script_name="search_debug.py", ...)

Database Interface

The database interface provides methods for building and managing vector databases. Currently, it provides a simple interface that can be extended:

use_skill(
  skill_name="rag",
  action="run_script",
  script_name="database.py",
  script_args="--action build --output_dir ./vectorstore --model evomaster/skills/rag/local_models/all-mpnet-base-v2"
)

Configuration

Model Selection

Default model: evomaster/skills/rag/local_models/all-mpnet-base-v2 (768 dims)

Choose embedding model based on your needs:

• •Default (local): evomaster/skills/rag/local_models/all-mpnet-base-v2 (768 dims)
  • •Fast, no download needed
  • •Used with node_vectorstore/ (768 dims)
• •High accuracy (OpenAI): text-embedding-3-large with configurable dimensions (512, 768, 1024, 2048, 3072)
  • •Used with corresponding node_vectorstore_<dim>/ directories
  • •Requires API access
• •Accuracy priority (HuggingFace): sentence-transformers/all-mpnet-base-v2 (768 dims) - HuggingFace model
• •Speed priority: sentence-transformers/all-MiniLM-L6-v2 (384 dims) - Faster, smaller
• •Custom models: Any HuggingFace AutoModel-compatible model or local model path

Important: Always use the vectorstore directory that matches the embedding model and dimension used during indexing.

Distance Thresholds

Distance thresholds depend on the similarity metric:

• •L2 distance: Lower is better (typically < 1.0 for similar items)
• •Cosine distance: Lower is better (typically < 0.3 for similar items)
• •Inner product: Higher is better (typically > 0.7 for similar items)

FAISS default is L2 distance. Adjust thresholds accordingly.

Best Practices

1. •Chunk Size: Balance context (larger) vs specificity (smaller) - typically 500-1000 tokens
2. •Overlap: Use 10-20% overlap to preserve context at boundaries
3. •Metadata: Include source, timestamp, and other metadata in nodes_data.json
4. •Distance Thresholds: Set appropriate thresholds to filter irrelevant results
5. •Top-K Selection: Start with k=5-10, adjust based on recall needs
6. •Model Selection: Use larger models for accuracy, smaller for speed

Common Issues

• •Poor Retrieval: Check embedding quality, ensure proper encoding
• •Irrelevant Results: Adjust distance thresholds, check query formulation
• •Missing Information: Ensure documents are properly indexed
• •Slow Queries: Use smaller embedding models, optimize FAISS index
• •Memory Issues: Use CPU mode, reduce batch sizes

Integration with agentic4mle

This RAG implementation is compatible with the agentic4mle project's vectorstore structure. The actual database is located at evomaster/skills/rag/MLE_DATABASE/:

use_skill(
  skill_name="rag",
  action="run_script",
  script_name="search.py",
  script_args="--vec_dir evomaster/skills/rag/MLE_DATABASE/node_vectorstore/improve --nodes_data evomaster/skills/rag/MLE_DATABASE/node_vectorstore/improve/nodes_for_improve.json --query \"...\""
)

Next Steps

• •For advanced retrieval patterns, see reference.md
• •For database building, see database.py interface
• •For troubleshooting, check logs and distance metrics