Skill

LightRAG local ollama graphrag

Turnkey local GraphRAG: doc corpus -> KG -> multi-mode retrieval, no paid API.

Primitives inside (11)

• dedicated-db-per-autobuilt-graph discipline

  Point auto-KG builders (LightRAG etc.) at a DEDICATED Neo4j database, never a shared application DB - auto-generated schemas conflict with and pollute application graphs.

  When: Wiring an automatic graph-building tool to an existing Neo4j instance.

  Register interest — 1 credit (≈ €0.10) — no payment

• embeddingfunc-no-double-wrap gotcha-fix

  LightRAG functions already decorated with @wrap_embedding_func_with_attrs (ollama_embed is) must be called via .func — wrapping them AGAIN inside EmbeddingFunc() is a documented upstream error pattern; the lambda-inside-EmbeddingFunc idiom is the aging form.

  When: Wiring a custom or provider embedding function into LightRAG.

  Register interest — 1 credit (≈ €0.10) — no payment

• guarded-import-actionable-exit failsafe

  A bundled script whose heavy dependency may be absent must catch the ImportError and exit 1 with the exact install command — never a raw traceback — and the selftest asserts this guard path when the dep is genuinely missing.

  When: Shipping runnable scripts inside skills/packages where the target machine may lack the main dependency.

  Register interest — 1 credit (≈ €0.10) — no payment

• lightrag-32k-ctx-silent-truncation gotcha-fix

  LightRAG's entity-extraction prompts require >=32k effective context; a smaller Ollama ctx SILENTLY truncates extraction — the KG builds with missing entities and NO error — so pin num_ctx 32768 before any ingest.

  When: Any LightRAG ingest against an Ollama-served LLM.

  Register interest — 1 credit (≈ €0.10) — no payment

• lightrag-query-mode-selector method

  Choose LightRAG query mode per question type: naive = quick factual lookup, local = specific entities/relations, global = cross-document themes, hybrid = both, mix (2026 default) = general purpose.

  When: Issuing a query against a LightRAG-built knowledge graph.

  Register interest — 1 credit (≈ €0.10) — no payment

• ollama-default-ctx-varies-verify-dont-assume calibration

  Never state 'the Ollama default context is N' — 0.23.x defaults OLLAMA_CONTEXT_LENGTH to 4k/32k/256k DEPENDING ON VRAM (and upstream LightRAG cites 8k); the once-true '2048 default' aged into a wrong claim — verify per box and pin explicitly.

  When: Any doc or debugging step that reasons about Ollama's effective context window.

  Register interest — 1 credit (≈ €0.10) — no payment

• ollama-list-vs-ps-diagnosis gotcha-fix

  ollama list shows INSTALLED models, ollama ps shows RUNNING ones - to load a pulled model, ollama run <m> or just issue an API call against it.

  When: Diagnosing why an Ollama model appears absent or inactive.

  Register interest — 1 credit (≈ €0.10) — no payment

• pdf-preextract-before-text-ingest gotcha-fix

  Text-ingest APIs (rag.insert) take plain text — PDFs fail or garbage-ingest unless extracted first with pdfminer.six/pypdf; the ingest script only accepts pre-extracted text.

  When: Feeding a mixed corpus (PDFs + text/markdown) into a text-based ingest pipeline.

  Register interest — 1 credit (≈ €0.10) — no payment

• pip-vs-clone-examples-path-gap gotcha-fix

  Repo examples/ scripts (visualization helpers etc.) exist ONLY in a cloned repo — pip-installed users don't have them; docs that reference examples/<file>.py by relative path must flag the clone-only prerequisite.

  When: Documenting or following a workflow that mixes 'pip install <pkg>' with repo-relative example paths.

  Register interest — 1 credit (≈ €0.10) — no payment

• purge-derived-caches-after-config-fix gotcha-fix

  After fixing an upstream config or extraction bug (e.g. a truncated context window), delete the pipeline's derived caches / working dir before the re-run that validates the fix — LLM response caches (LightRAG's llm_response_cache.json in the SAME working_dir) replay the old broken outputs, so the fix appears not to work.

  When: Re-running an LLM pipeline (LightRAG ingestion or any cached extraction) after a config or code change that should alter previously-cached LLM responses.

  Register interest — 1 credit (≈ €0.10) — no payment

• research-batches-until-dry discipline

  Research by aggregation: read REAL sources in batches, fold in only genuinely NEW findings per batch, and stop only when a full batch adds zero new items - then declare the research converged/dry.

  When: Grounding or lifting an artifact against a body of sources where coverage completeness matters and sampling bias must be avoided.

  Register interest — 1 credit (≈ €0.10) — no payment

Neighbour skills

Skills whose primitives are closest to this one (bge-m3 similarity):

• Instructor pydantic extractor
• Agentic design pattern selector
• DSPy system prompt compiler