Skip to content

Dashboard Walkthrough

The AutoChunks Dashboard is your command center for understanding precisely why a chunking strategy succeeds or fails.

1. Setup and Configuration

Upon launching autochunks serve, you are greeted by the Control Center.

  • Documents Path: Point this to your document repository (supports PDF, MD, HTML, TXT).
  • Optimization Mode:
    • Light (Fast Sample Sweep): Optimized for development. It caps document sampling to find the winning strategy quickly.
    • Full (Exhaustive Corpus Audit): Production-grade tuning. Uses a wider document sample for high-confidence benchmarking.
  • Embedding Provider:
    • Local (Production Simulation): Uses real vector embeddings via the sentence-transformers library. This allows you to match your production environment (e.g., BGE-Large).
    • OpenAI: Uses the OpenAI Embeddings API. Requires an API key which is entered securely in the dashboard.
    • Ollama: Connects to a locally running Ollama instance.
    • Hashing (Architectural/CI): A high-performance mock provider used for structural testing and CI/CD integration without model overhead.

2. The Sweep

When you click Start Optimization, the engine enters the tournament state.

  1. Ingestion: Files are read and indexed via SHA-256 fingerprinting.
  2. QA Generation: Synthetic ground-truth questions are created for evaluation.
  3. Parallel Evaluation: The system fans out to test all selected strategies simultaneously across the vectorized search pipeline.

3. The Leaderboard

Once complete, the dashboard displays the Winning Strategy and a ranked list of all candidates.

  • Objective Score: The weighted composite score used to determine the winner based on your Optimization Goal.
  • nDCG / MRR: Detailed retrieval benchmarks and ranking positions.
  • Chunk Count: Analyzes the fragmentation overhead; some strategies may score highly but produce an excessive number of chunks, increasing storage costs.

4. Fidelity Inspector

This tools is designed for qualitative debugging. Click Compare on any candidate to open the Inspector.

  • Side-by-Side View: See exactly where the Winner cut the text vs. any other candidate strategy.
  • Visual Diff: Identifies structural errors, such as whether a sentence was bifurcated or if a semantic splitter triggered correctly.
  • Sync Scroll: Synchronized scrolling across panes to track long-form documents.

5. Deploy as Code

The Code Export tab provides framework-specific snippets for LangChain, Haystack, or LlamaIndex. Copy these configurations to instantly apply the optimized strategy to your production data pipeline.