Skip to content

v1.0.0-beta.11

Choose a tag to compare

View all tags
@KikeVen KikeVen released this 28 Jun 22:59
· 18 commits to main since this release
v1.0.0-beta.11
707b39f

semantic reliability update

I have implemented the Strict Attribution framework to eliminate "Semantic Smearing" and ensure the assistant remains grounded in the provided codebase context.

1. Key Improvements

  • Strict Attribution Layer: Added explicit Grounding, Signature Truth, and No Hallucination rules to both DeepSeek (Cloud) and Mistral (Local) synthesis prompts in main.py.
  • Automated Audit Pipeline: Created reliability_audit.py to programmatically compare Ground Truth (ChromaDB raw data) against LLM responses.
  • Local Model Upgrade: Transitioned to mistral:7b as the default local engine for better instruction adherence and complex logic analysis.
  • Deterministic Routing: Verified that the tree-sitter architecture in code_indexer.py correctly isolates entity signatures, preventing logic from "bleeding" between adjacent functions.

2. Verified Results

Phase 2 audits confirm that models now explicitly state when information is missing and correctly attribute implementation details only to the specific functions where they are defined.

Upper-case constant indexing

  • Making UPPER_CASE constants searchable — before my changes, tree-sitter would skip config.py entirely because it had no functions/classes (it was a "bare" file). Now the indexer extracts constants like FETCH_CAP, QUERY_DISTANCE_THRESHOLD, DEEPSEEK_PRICING, etc., with their preceding comments as searchable context.
  • JS/TS constant extraction — same for JavaScript/TypeScript files.
  • Checkbox indexing — todo.md checkboxes are now individual entities, so you can query "pending tasks".

brief generation

Converted all 9 brief section ChromaDB queries from keyword soup to natural language for better semantic matching with the embedding model. Updated prompt templates to use generic, context-driven categories — removing
assumptions like "Frontend" and "Backend" that invite hallucination, and adding omit-if-not-present instructions so the LLM reports only what the codebase actually contains.

embedding-docstring skill

  • Walks the workspace directory tree
  • Loads .memignore once and builds an exclusion list
  • Produces a file inventory (filtered) before any entity inventory begins
  • Then chains into the existing STEP 1–3 per file, with the 10-entity cap applying per file
  • Fixed 300 → 400 in flag example
  • "What you check" — added instruction to scan #/// blocks above UPPER_CASE constants
  • "What you fix" — added constant reading step
  • "Applies to" — added walk-backward rules for how constant comment blocks are found

This file is important to run before any other entity inventory, as it will produce a file inventory that is used to filter out files that should not be processed in the subsequent steps.

Assets 2
Loading