- Retitle to reflect actual scope: taxonomy, contract, classifier strategy
- Split Decision section into permanent (taxonomy, model assignment) vs
tactical [BOOTSTRAP] (LLM classifier)
- Mark LLM classifier explicitly as interim implementation with pointer
to Future Path
- Clarify that Routing Contract is implementation-independent
- Consolidate prompt engineering rules as symptoms of architectural mismatch
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add PLATFORM query type that bypasses RAG and uses a lighter model
- Introduce OLLAMA_MODEL_NAME_CONVERSATIONAL env var to route CONVERSATIONAL
and PLATFORM queries to a separate (smaller) Ollama model
- Replace raw message history in classifier with compact intent history
(classify_history) to eliminate anchoring bias in small models
- Add <history_rule> and <platform_priority_rule> to classifier prompt so
the model evaluates each message independently while still resolving
ambiguous references from prior turns
- Add fast-path detection for known platform-injected prompt prefixes
- Add PLATFORM_PROMPT for account/metrics/usage responses
- Persist classify_history in classify_history_store alongside session_store
- Document decisions in ADR-0008
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Created `n00 Beir Analysis_cosqa.ipynb` for analyzing CoSQA dataset with BEIR.
- Created `n00 first Analysis.ipynb` for initial analysis using Ragas and Ollama embeddings.
- Implemented `evaluate_embeddings_pipeline.py` to evaluate embedding models across CodexGlue, CoSQA, and SciFact benchmarks.
- Added adapters for Ollama and HuggingFace embeddings to ensure compatibility with BEIR.
- Included functions to load datasets and evaluate models with detailed metrics.
rafa-ruiz
izapata
acano