Add raw request/response logging with RAGAS_ENABLE_HTTP_LOG env flag

fixes: #1907

…#1906)

#1904

The issue arises because `self.single_turn_prompt.instruction` and
`self.multi_turn_prompt.instruction` were not being properly assigned
during the initialization of `SimpleCriteriaScore()`.

This bug probably raise a KeyError since `personas` is not set.

This could be a bug due to an oversight.

Related to #1917

…r) (#1910)

This PR updates the documentation to correctly describe the Semantic
similarity.

### Issue
The documentation previously stated that a **cross-encoder** was used
for computing the semantic similarity score. However, after reviewing
the implementation, it is clear that the current approach follows a
**bi-encoder** strategy:

- The ground truth and response are encoded independently
- Their embeddings are then compared using cosine similarity

A cross-encoder would typically process both texts together in a single
forward pass (e.g., concatenating them before encoding), which is not
the case in the current implementation.

### Current Implementation
For example, in the current implementation:

```python
embedding_1 = np.array(await self.embeddings.embed_text(ground_truth))
embedding_2 = np.array(await self.embeddings.embed_text(answer))
# Normalization factors of the above embeddings
norms_1 = np.linalg.norm(embedding_1, keepdims=True)
norms_2 = np.linalg.norm(embedding_2, keepdims=True)
embedding_1_normalized = embedding_1 / norms_1
embedding_2_normalized = embedding_2 / norms_2
similarity = embedding_1_normalized @ embedding_2_normalized.T
score = similarity.flatten()
```

This code shows that the ground truth and response are encoded
separately, and their similarity is computed using cosine similarity,
which is characteristic of a **bi-encoder** approach.

### Fix
The term "cross-encoder" has been corrected to "bi-encoder" in the
documentation to ensure consistency with the actual implementation.

first iteration of Nvidia accuracy metric
Added a simple implementation for single turn accuracy,
context_relevance and answer_groundedness metrics

---------

Co-authored-by: jjmachan <[email protected]>

Add @t.runtime_checkable to the ModeMetric protocol to allow runtime
type checking using isinstance() and issubclass().

- r2r integration
- fixed spelling mistake in swarm tutorial
- added @t.runtime_checkable to metric mode

---------

Co-authored-by: Jithin James <[email protected]>

Co-authored-by: Jithin James <[email protected]>

Easy fix Groundedness metric, 5 retries early break.
Added logger to the 3 nv_metrics retries.
Fixed input context max lenght to 7k to avoid 8k break.

- the current code was logging a name and mode of noise sensitivity
metric instead of just the name

When saving the knowledge graph, all the fields of each node is no
longer stored in the relationships.

This can save a very large amount of hard disk space.

Use json.dumps() on the dict output of Pydantics's
BaseModel.model_json_schema() to ensure the string representation is
valid JSON using double quotes instead of single quotes.

…th Token Count 101–500 (#1934)

The `default_ transforms` function defined at
`src/ragas/testset/transforms/default.py` has a problem with handling
transforms for documents with 101-500 tokens.

The code divides the `transforms` configurations based on the document's
token count. Several transforms are instantiated when the ["101-500"
token count bins the first quartile
(Q1](https://github.com/explodinggradients/ragas/blob/2bc29a2b8358ddb6b167fdf7ab0518ad9371463c/src/ragas/testset/transforms/default.py#L128),
among them the `cosine_sim_builder`. While `cosine_sim_builder` is
correctly **instantiated** (line 139), it's then **not included** in the
list of transforms that are actually returned (line 153).

It appears that `cosine_sim_builder` was likely unintentionally omitted
from the returned transforms list. The intended behavior should probably
mirror how `ner_overlap_sim` is handled (line 120), where
`cosine_sim_builder` is instantiated and added to the returned list. The
current code effectively instantiates `cosine_sim_builder` but then
discards it. This omission might impact the number of relationships
created in the knowledge graph.

1. I have noticed that if `response.statements` is an empty list, then
the output of `verify_claims` function is `array([], dtype=float64)`,
which raises a type error when `~` operation is applied to
`reference_response` variable. This occurs in a rare case when
hypothesis_list is empty (when there are no claims in the response).
2. A small change is the choice of words in `LLMDidNotFinishException`

---------

Co-authored-by: Mikhail Zybin <[email protected]>