| --- |
| license: mit |
| tags: |
| - langchain |
| - rag |
| - bayesian |
| - probabilistic |
| - uncertainty |
| - retrieval |
| --- |
| |
| # Bayesian LangChain RAG |
|
|
| A corrected, Hugging Face-ready implementation of uncertainty-aware retrieval augmented generation built on **LangChain Core**. |
|
|
| This repository replaces the earlier heuristic "tool succeeded = high likelihood" and beam-style scoring with an explicit generalized-Bayesian retrieval posterior. For a query `q` and candidate document `d`, |
|
|
| `p(d | q) ∝ p(d) (similarity(q,d) + eps)^tau` |
|
|
| where the prior `p(d)` is built from a Beta source-reliability mean, source authority, and freshness metadata. The similarity term is deliberately described as a **generalized-Bayes/Gibbs likelihood**, not as a literal language-model likelihood. |
|
|
| ## What was corrected |
|
|
| - Removed notebook/Colab shell commands and hard-coded paths from library code. |
| - Removed the non-Bayesian `p_langchain` beam scorer from the Bayesian runtime. |
| - Replaced deterministic identical-particle behavior with an importance-corrected SMC utility for workflows that need particles. |
| - Added proposal-density correction `log target - log q` to SMC weights. |
| - Fixed all-`-inf` log-weight normalization. |
| - Preserved final posterior weights by forbidding final-step resampling. |
| - Added a real `BaseRetriever` implementation and LangChain `Runnable` entry point. |
| - RAG now reasons over retrieved evidence quality rather than rewarding any successful tool call. |
| - Added explicit abstention for missing evidence and unresolved near-equal conflicts. |
| - Added citation-grounded answer synthesis and a reproducible end-to-end RAG benchmark. |
|
|
| ## Install |
|
|
| ```bash |
| pip install -e . |
| ``` |
|
|
| No GPU is required. The default evaluation generator is extractive and grounded. Any LangChain-compatible model can be supplied for generative answering. |
|
|
| ## Quick start |
|
|
| ```python |
| from bayesian_langchain import DocumentRecord, BayesianIndex, BayesianRAG |
| |
| docs = [ |
| DocumentRecord( |
| id="official", |
| title="Current portal policy", |
| text="The current HR portal is PeopleHub.", |
| source="official", |
| metadata={"reliability_alpha": 18, "reliability_beta": 2, "authority": 0.98, "freshness": 1.0}, |
| ), |
| DocumentRecord( |
| id="legacy", |
| title="HR portal guide", |
| text="The HR portal is WorkSphere.", |
| source="legacy-cache", |
| metadata={"reliability_alpha": 6, "reliability_beta": 6, "authority": 0.45, "freshness": 0.18}, |
| ), |
| ] |
| |
| rag = BayesianRAG(BayesianIndex(docs)) |
| print(rag.invoke("What is the current HR portal?")) |
| ``` |
|
|
| `BayesianRAG.as_runnable()` returns a standard LangChain `RunnableLambda`. |
|
|
| ## RAG validation |
|
|
| Run: |
|
|
| ```bash |
| PYTHONPATH=src:. python evaluation/rag_benchmark.py |
| ``` |
|
|
| The committed result is in `artifacts/rag_benchmark_results.json`. |
|
|
| | Suite | System | Recall@1 | MRR | Answer accuracy | Abstention accuracy | End-to-end accuracy | |
| |---|---|---:|---:|---:|---:|---:| |
| | Clean RAG (8 QA) | Bayesian | 1.000 | 1.000 | 1.000 | n/a | 1.000 | |
| | Clean RAG (8 QA) | Plain TF-IDF | 1.000 | 1.000 | 1.000 | n/a | 1.000 | |
| | Conflict/missing RAG (18 QA) | Bayesian | 1.000 | 1.000 | 1.000 | 1.000 | 1.000 | |
| | Conflict/missing RAG (18 QA) | Plain TF-IDF | 0.250 | 0.611 | 0.250 | 0.333 | 0.278 | |
|
|
| The conflict suite deliberately contains stale high-overlap documents, current authoritative documents, questions with no supporting evidence, and equal-authority contradictions. The results are **controlled repository tests**, not a claim of universal RAG accuracy. |
|
|
| ### Framework-level LangChain RAG test |
|
|
| The repository also contains `evaluation/framework_rag_benchmark.py`, which routes every case through the real LangChain path: |
|
|
| `BayesianRetriever(BaseRetriever) -> BayesianRAG.as_runnable() -> grounded generation/abstention`. |
|
|
| It was validated against **langchain-core 1.6.11**. Results: **8/8 clean cases** and **18/18 conflict/missing-evidence cases** passed end-to-end. The committed output is `artifacts/framework_rag_benchmark_results.json`. |
|
|
| Run it with: |
|
|
| ```bash |
| PYTHONPATH=src:. python evaluation/framework_rag_benchmark.py |
| ``` |
|
|
| ## Layout |
|
|
| - `src/bayesian_langchain/core.py` — posterior retrieval, abstention, grounded answerer |
| - `src/bayesian_langchain/retriever.py` — LangChain `BaseRetriever` |
| - `src/bayesian_langchain/rag.py` — end-to-end LangChain RAG and Runnable |
| - `src/bayesian_langchain/smc.py` — corrected importance-weighted SMC utility |
| - `evaluation/rag_benchmark.py` — controlled retrieval/grounding benchmark |
| - `evaluation/framework_rag_benchmark.py` — real LangChain `BaseRetriever`/Runnable RAG benchmark |
| - `tests/` — regression and integration tests |
| - `VALIDATION.md` — exact framework/version validation record |
|
|
