pydocs-mcp 0.3.0

Published to PyPI: https://pypi.org/project/pydocs-mcp/0.3.0/

pip install pydocs-mcp

Added (LLM tree-reasoning — enrichment, token budget, two-stage rerank)

• PageIndex node enrichment — each LLM-visible tree node now carries its real
  signature (params + type hints + return annotation), its decorators, and a
  docstring excerpt, beyond the generated summary. Tunable via doc_excerpt
  (sections | full | off) and doc_excerpt_max_chars. A non-destructive
  schema auto-refresh (v9) re-extracts the metadata on next index without
  re-embedding unchanged chunks.
• Token-counted tree budget — the serialized tree handed to the LLM is bounded
  in real tiktoken tokens (previously whitespace words, which under-counted code
  ~3× and could overflow the model's context window with a 400
  context_length_exceeded). max_tree_words → max_tree_tokens
  (int | None; None auto-derives from the configured model's context window).
  Over-budget pruning is content-first — drop per-node doc excerpts before whole
  nodes. Adds tiktoken as a runtime dependency.
• BM25 → tree two-stage rerank — opt-in rerank_candidates mode on the
  llm_tree_reasoning step scopes the LLM-visible tree to a prior BM25/dense
  candidate set and writes its ranked picks back as the pipeline's final ranking
  (with a repoqa_bm25_tree_rerank benchmark config).
• Persist chunks.qualified_name (schema v7) so tree-reasoning picks resolve to
  the correct chunks.

Added (on-device dense embeddings)

• sentence_transformers embedding provider (provider: sentence_transformers)
  serving Qwen/Qwen3-Embedding-0.6B and other SentenceTransformer models via
  torch — a GPU-reliable on-device dense embedder (torch frees CUDA memory
  between sequential index-builds). Opt-in via the [sentence-transformers]
  extra. New EmbeddingConfig knobs max_seq_length / normalize /
  query_prompt_name (the first two fold into the pipeline hash; the
  query-only prompt does not).

Removed

• The onnx embedding provider (OnnxEmbedder and the onnx_file /
  query_instruction config fields). The torch-backed sentence_transformers
  provider replaces it for on-device Qwen3-Embedding — onnxruntime leaked the
  CUDA arena across the benchmark's sequential index-builds.

Added (GPU inference)

• --gpu flag on serve, index, and watch (and the benchmark runner)
  to run all embedder inference — FastEmbed, the sentence_transformers
  provider, and PyLate late-interaction — on CUDA. No YAML change; covers both
  index-time and query-time embedding. The execution device is excluded from the
  pipeline / index-cache hash, so toggling --gpu shares the same .tq /
  fast-plaid index and never forces a re-index (it is a latency knob, not a
  quality change).
• EmbeddingConfig.device (cpu / cuda) wiring through build_embedder
  into the FastEmbed and sentence_transformers embedders;
  AppConfig.with_device(gpu=...) stamps the device after config load. GPU
  runtimes (onnxruntime-gpu, fastembed-gpu, CUDA torch) are documented in
  INSTALL.md, not auto-installed.