v2.6.0

Added

• OpenAIConnector baseURL for OpenAI-compatible APIs (Groq,
  Together, the Hugging Face router, vLLM, LM Studio, Ollama's /v1
  endpoint). With a custom baseURL the key is never read from
  OPENAI_API_KEY, so an OpenAI key is not sent to another provider; pass
  apiKey, or omit it for local servers that need none.
• validatePlugin(stage, plugin, { registry }) returns
  { valid, errors, warnings } for a plugin without throwing, for use in
  plugin tests. registry: true also applies the rules
  pluginRegistry.register() enforces.
• ParallelEmbedder mode ('auto', 'batch', 'single').
• checkAnswer({ question, answer, sources }) checks an answer from any
  framework (LangChain, LlamaIndex, Mastra, the AI SDK, your own code)
  against its sources: grounded, groundedness, the unsupported
  sentences, per-sentence citations and evaluation scores. Sources may be
  strings, { id, content }, LangChain Documents or LlamaIndex nodes;
  normalizeSources is exported for the conversion alone.
• Caching: createRagPipeline({ cache }) caches query embeddings,
  retrieval results and answers (true, a store, or per layer with time to
  live). MemoryCache (LRU and TTL) and RedisCache (node-redis or
  ioredis) are included; any get/set store works. A layer caches only
  when its component identifies itself (getCacheKey(role) or cacheKey;
  all built-ins do), retrieval is keyed by the store's state and the query
  vector, answers by the exact documents and generation options, and
  answers are not cached when the temperature is above 0. Retrieval is
  cached before filterResults and the reranker, so guardrail ACLs run on
  every request. Nothing from a failed, aborted or timed-out run, or from an
  unfinished stream, is stored. Results carry cache per layer and the
  pipeline has cacheStats().
• @devilsdev/rag-pipeline-utils/testing: deterministic, offline test
  helpers. HashEmbedder, EchoLLM (answers with the top document or a set
  answer, records every call, streams), fixtures.documents and
  createTestPipeline(). Shares classes with the main entry and is not
  loaded by it.
• computeSemanticSimilarity(answer, reference, { embedder }): cosine
  similarity of the two embeddings, 0 to 1. checkAnswer adds it as
  evaluation.scores.semanticSimilarity when given a referenceAnswer and
  an embedder.
• rag-pipeline query options: --trace (time per stage), --explain
  (passage scores, and the passages supporting each answer sentence, with
  unsupported sentences flagged), --stats (latency, estimated tokens and
  cost), --streaming and --timeout <ms>; global --quiet. --json
  includes trace, stats and explanation when asked.
• createAiSdkAdapter({ ai, embeddingModel, languageModel }) wraps
  Vercel AI SDK models as an embedder and an llm for
  createRagPipeline(), with streaming, embedMany batching and the run's
  AbortSignal. You pass the AI SDK functions, so it is not a dependency and
  works with AI SDK 5, 6 and 7 from ESM or CommonJS. The RAG system prompt
  is sent as both instructions (AI SDK 7) and system (5 and 6). Tested
  against AI SDK 6 and 7 with their mock models.

Changed

• Faster CLI. Commands load only their own dependencies: --help
  takes about 0.5 s and query about 0.5 s instead of about 3.6 s. Loading
  the CLI drops from 1281 modules and 93 packages to 55 and 13. The HTML,
  Markdown and CSV loaders load jsdom, marked and csv-parse when they
  read a file.
• Faster import. require('@devilsdev/rag-pipeline-utils') takes about
  125 ms instead of about 515 ms and loads 31 modules from 9 packages
  instead of 245 from 54 (published bundle, Node 22). axios,
  jsonwebtoken, ajv and chokidar now load when the feature that needs
  them is first used (the Anthropic, Cohere and Ollama connectors,
  JWTValidator, plugin registration and config validation, hot reload).
  The shared pluginRegistry loads its contracts on first use; registries
  created with new PluginRegistry() still load them in the constructor.

Fixed

• ExecutionTracer.exportTraces('json') threw "circular structure"
  with the default captureOutputs: true: traces stored the live result,
  which carries the trace. Traces now store JSON snapshots, which also keeps
  later changes to a result out of its recorded trace.
• ParallelEmbedder failed with the built-in connectors, which embed
  one string at a time, because it always called embed(batch) with an
  array. In the default 'auto' mode it now embeds texts one by one for
  those connectors and uses embedBatch(texts) when an embedder has it;
  other embedders are still called with the array.