Repository navigation
Releases
Releases · anjanatiha/Retrieval-Augmented-Generation-RAG-Agent
v1.4.5
What's Changed
- Added commercial standard principle to ARCHITECT_PROMPT.md — applies to code, docs, repo, UI, and tests
- Fixed test count (848 → 859) across CONTRIBUTING.md
- Updated wiki Home with full feature list, XLS format, test counts, Docker install link
- Updated wiki Contributing with test counts and installation guide link
- Closed Dependabot PR — gradio intentionally pinned to 5.23.0
No functional changes — documentation and repository hygiene.
v1.4.4
What's Changed
- Added SECURITY.md — vulnerability reporting policy
- Added GitHub issue templates (Bug Report, Feature Request) and PR template
- Added CHANGELOG.md — full version history from v1.0.0
- Updated README, CONTRIBUTING, wiki with CHANGELOG and SECURITY references
- Fixed test count (848 → 859 local), added XLS to 9-format list
- Updated wiki Home with all features, test counts, Docker install link
- Updated release badge to v1.4.4
No functional changes — documentation and repository hygiene.
v1.4.3
What's Changed
- Added security policy (SECURITY.md) with vulnerability reporting instructions
- Added GitHub issue templates (Bug Report, Feature Request)
- Added pull request template with checklist
No functional changes — documentation and repository hygiene only.
v1.4.2
What's New
Multi-Turn Conversation
- The chatbot now maintains context across turns within a session
- Follow-up questions work correctly — pronouns ("she", "it"), references ("that document", "the person mentioned earlier"), and continuation phrases ("tell me more about that") are resolved from prior conversation turns
- All factual claims remain grounded in retrieved document context — the anti-hallucination constraint is preserved
Docker Support
- Run the full stack with a single command — no Python setup required
docker compose upstarts the app and Ollama as separate containers- ChromaDB and Ollama models persisted in named Docker volumes
- NVIDIA GPU support via
docker-compose.gpu.yml
Tests
859 local · 385 HF Space · 1,244 total — all passing
v1.4.1
What's New
Docker Support
- Run the full stack with a single command — no Python setup required
docker compose upstarts both the app and Ollama as separate containers- ChromaDB and Ollama models persisted in named Docker volumes — no re-indexing or re-downloading on restart
./docs/mounted as a bind mount — add documents without rebuilding the image- NVIDIA GPU support via
docker-compose.gpu.ymloverride
# Pull models once (~2.3 GB)
docker compose run --rm ollama ollama pull hf.co/CompendiumLabs/bge-base-en-v1.5-gguf
docker compose run --rm ollama ollama pull hf.co/bartowski/Llama-3.2-3B-Instruct-GGUF
# Start
docker compose upOpen http://localhost:8501.
Tests
859 local · 385 HF Space · 1,244 total — all passing
v1.4.0
What's New
RAGAS Evaluation
- New
python main.py --ragascommand runs four industry-standard LLM-as-judge metrics - Faithfulness — every claim grounded in retrieved context
- ResponseRelevancy — answer addresses the question asked
- ContextPrecision — best-matching chunks ranked highest
- ContextRecall — retrieved context covers the ground truth
- Uses the same test cases as
--benchmark— results are directly comparable - Optional dependencies:
pip install -e ".[eval]"— app runs normally without them
Pre-Commit Hooks
- Automated code quality checks run before every
git commit - Hooks: trailing whitespace · end-of-file · YAML/JSON/TOML validation · isort · flake8 (syntax errors only)
- Install with:
pip install -e ".[dev]"thenpre-commit install
Documentation
- README, BENCHMARK.md, and all wiki pages updated to reflect new features
- RAGAS evaluation section added to docs_technical/BENCHMARK.md
Tests
859 local · 385 HF Space · 1,244 total (+16 RAGAS evaluation tests)
v1.3.2
What's New
Streaming Responses
- Answers now appear word by word as they are generated — no waiting for the full response
- Progress bar clears at the start of generation; the live token stream is the feedback
- Terminal (CLI) behaviour unchanged
Engineering
- Pipeline split into prepare and finalize phases to support streaming
- Query classification, retrieval budgeting, and prompt building extracted into a dedicated module
- All files within the 500-line limit
Tests
- 843 local · 385 HF · 1,228 total — 0 failures
Full Changelog
v1.3.1
What's Changed
- Added type hints to all method signatures across
src/rag/— every parameter and return type is now explicit - Expanded docstrings to Google style on all public and non-obvious private methods
- No logic changes — all existing functionality preserved
Test Results
- Local: 843 passed
- HF Space: 385 passed, 2 skipped
- Total: 1,228 passed
Full Changelog
v1.3.0
What's New
Features
- Translate tool — agent tool #6, any-to-any language translation; short queries search the knowledge base first
- Topic search — search DuckDuckGo and auto-index top results, no API key required
- Recursive URL crawl — configurable depth, page budget, and keyword filter
- Clear Added Content — one-click reset for runtime-ingested documents
- URL scheme auto-correction — bare domains are automatically prefixed with
https://
Quality
| Metric | Before | After |
|---|---|---|
| Overall benchmark | 0.789 | 0.808 ▲ |
| Faithfulness | 0.802 | 0.967 ▲ |
| MRR | 0.900 | 1.000 ▲ |
| Tool benchmark | 12/12 | 18/18 ▲ |
Engineering
- 500-line file size limit enforced — 6 new modules extracted
- Tests: 807 → 848 local, 344 → 385 HF, 1,233 total
Full Changelog
v1.2.2
What's New
Features
- Translate tool — translate to any language; short queries search the knowledge base first
- Modular reranker — type-aware reranking extracted into its own module
Fixes
- 19 CI test failures resolved
- Entry point slimmed to under 50 lines
Tests
- 843 local · 387 HF · 1,230 total — 0 failures