graphgrail 0.1.0__tar.gz

graphgrail-0.1.0/.gitignore

@@ -0,0 +1,62 @@
+# --- Legacy/proprietary import baseline ---
+# The _legacy_source/ tree is a verbatim copy of the proprietary `nirvana_agents/utilities/graphrag`
+# snapshot used as the starting point for the GRAIL migration. It is intentionally kept in-tree
+# so Claude Code sessions are self-contained (no path dependency on the proprietary monorepo).
+# It must NEVER be committed to the public GRAIL repo. Delete the directory once migration is complete.
+_legacy_source/
+
+# --- Python ---
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+*.egg
+*.egg-info/
+build/
+/dist/
+.eggs/
+pip-wheel-metadata/
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.tox/
+.coverage
+.coverage.*
+htmlcov/
+.hypothesis/
+
+# --- Virtual envs ---
+.venv/
+venv/
+env/
+ENV/
+
+# --- Node (frontend builds) ---
+node_modules/
+
+# --- IDE / OS ---
+.idea/
+.vscode/
+.DS_Store
+Thumbs.db
+
+# --- GRAIL runtime artifacts (default working dir) ---
+# Outputs of indexing pipelines (parquet, lancedb, graphml). User overrides via config.
+output/
+cache/
+lancedb/
+*.parquet
+*.graphml
+mapping.json
+
+# --- Notebook generated output ---
+notebooks/_pyvis_graph.html
+notebooks/.ipynb_checkpoints/
+
+# --- Secrets / env ---
+.env
+.env.*
+!.env.example
+*.pem
+*.key

graphgrail-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 Cámara Chilena de Inteligencia Artificial
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

graphgrail-0.1.0/PKG-INFO

@@ -0,0 +1,68 @@
+Metadata-Version: 2.4
+Name: graphgrail
+Version: 0.1.0
+Summary: Graph RAG with Advanced Integration and Learning — an open-source GraphRAG library with incremental updates, file-level provenance, and pluggable LLM/storage backends.
+Project-URL: Repository, https://github.com/CAMARA-CHILENA-INTELIGENCIA-ARTIFICIAL/GRAIL
+Author-email: Benjamin González Guerrero <ben@nirvana-ai.com>
+License: Apache-2.0
+License-File: LICENSE
+Keywords: graphrag,knowledge-graph,leiden,llm,rag,retrieval
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Requires-Dist: aiohttp>=3.9
+Requires-Dist: aiosqlite>=0.20
+Requires-Dist: graspologic>=3.4
+Requires-Dist: httpx>=0.27
+Requires-Dist: lancedb>=0.10
+Requires-Dist: networkx>=3.2
+Requires-Dist: numpy>=1.24
+Requires-Dist: openai<2,>=1.40
+Requires-Dist: pandas>=2.0
+Requires-Dist: pyarrow>=14
+Requires-Dist: pydantic<3,>=2.5
+Requires-Dist: pypdf>=5.0
+Requires-Dist: python-docx>=1.1
+Requires-Dist: python-dotenv>=1.0
+Requires-Dist: pyyaml>=6
+Requires-Dist: rich>=13
+Requires-Dist: scikit-learn>=1.3
+Requires-Dist: scipy>=1.10
+Requires-Dist: tenacity>=8.2
+Requires-Dist: textual>=0.50
+Requires-Dist: tiktoken>=0.7
+Requires-Dist: typer>=0.12
+Provides-Extra: chroma
+Requires-Dist: chromadb>=0.5; extra == 'chroma'
+Provides-Extra: dev
+Requires-Dist: mypy>=1.10; extra == 'dev'
+Requires-Dist: pytest-asyncio>=0.23; extra == 'dev'
+Requires-Dist: pytest-mock>=3.12; extra == 'dev'
+Requires-Dist: pytest>=8; extra == 'dev'
+Requires-Dist: ruff>=0.5; extra == 'dev'
+Provides-Extra: faiss
+Requires-Dist: faiss-cpu>=1.7; extra == 'faiss'
+Provides-Extra: notebook
+Requires-Dist: ipykernel>=6.29; extra == 'notebook'
+Requires-Dist: ipywidgets>=8.1; extra == 'notebook'
+Requires-Dist: jupyter>=1.0; extra == 'notebook'
+Requires-Dist: matplotlib>=3.8; extra == 'notebook'
+Requires-Dist: pyvis>=0.3.2; extra == 'notebook'
+Provides-Extra: rerank
+Requires-Dist: sentence-transformers>=2.7; extra == 'rerank'
+Provides-Extra: s3
+Requires-Dist: aioboto3>=12; extra == 's3'
+Requires-Dist: boto3>=1.34; extra == 's3'
+Provides-Extra: ui
+Requires-Dist: bcrypt==4.0.1; extra == 'ui'
+Requires-Dist: fastapi>=0.115; extra == 'ui'
+Requires-Dist: passlib[bcrypt]>=1.7; extra == 'ui'
+Requires-Dist: python-jose[cryptography]>=3.3; extra == 'ui'
+Requires-Dist: sse-starlette>=2.0; extra == 'ui'
+Requires-Dist: uvicorn[standard]>=0.30; extra == 'ui'

graphgrail-0.1.0/configs/community.yaml

@@ -0,0 +1,33 @@
+# Community detection + reporting config. See docs/community.md.
+
+# Hierarchical Leiden (graspologic).
+max_cluster_size: 50
+use_lcc: false
+strategy: leiden
+seed: null  # set an int for reproducible clustering
+
+# Per-stage endpoint+model overrides. When null, inherit from configs/llm.yaml.
+community_report_endpoint: null
+community_report_model: null
+json_corrector_endpoint: null
+json_corrector_model: null
+
+# JSON community report ceiling.
+max_report_length: 8192
+
+# Concurrency for community report LLM calls. null = use the global
+# llm.concurrent_requests limit. Set lower to avoid overwhelming the
+# endpoint with heavy report-generation calls.
+# report_concurrency: null
+
+# If true, include claims (covariates) in the report context.
+include_covariates: false
+
+# Incremental updates — change ratio threshold above which we re-cluster
+# the affected subgraph instead of label-propagating new nodes.
+incremental_change_threshold: 0.3
+
+# After Leiden, small communities below this size get merged with their nearest
+# centroid via DBSCAN over the entity embeddings.
+min_community_size: 10
+embedding_merge_eps: 0.5

graphgrail-0.1.0/configs/embeddings.yaml

@@ -0,0 +1,19 @@
+# Embedding client config. See docs/llm.md.
+#
+# Same shape as configs/llm.yaml. endpoint references entries in
+# configs/endpoints.yaml. The defaults below use DeepInfra's hosting of
+# intfloat/multilingual-e5-large — a strong open-weights multilingual embedder.
+
+endpoint: deepinfra
+model: intfloat/multilingual-e5-large
+encoding_format: float
+
+# Batching / concurrency
+max_batch_size: 1024
+concurrent_requests: 30
+
+# Timeouts / retries
+request_timeout: 180.0
+max_retries: 10
+max_retry_wait: 10.0
+sleep_on_rate_limit: 30.0

graphgrail-0.1.0/configs/endpoints.yaml

@@ -0,0 +1,76 @@
+# Endpoint definitions. See docs/llm.md.
+#
+# Every entry here is a deployment of the OpenAI Chat Completions / Embeddings
+# protocol. GRAIL is endpoint-agnostic by design — there are no vendor-specific
+# code paths. To use a different vLLM cluster, sglang server, regional OpenAI
+# proxy, or anything else, add an entry below and reference it by name in
+# llm.yaml / embeddings.yaml / per-stage overrides.
+#
+# These defaults match the built-in registry in grail/llm/providers.py. User
+# edits MERGE with the defaults (per-endpoint deep merge), so you only need to
+# include entries you actually want to override.
+
+openai:
+  base_url: https://api.openai.com/v1
+  api_key_env: OPENAI_API_KEY
+
+anthropic:
+  base_url: https://api.anthropic.com/v1
+  api_key_env: ANTHROPIC_API_KEY
+  notes: Anthropic exposes /v1/messages over the OpenAI-compatible protocol.
+
+deepinfra:
+  base_url: https://api.deepinfra.com/v1/openai
+  api_key_env: DEEPINFRA_API_KEY
+
+together:
+  base_url: https://api.together.xyz/v1
+  api_key_env: TOGETHER_API_KEY
+
+groq:
+  base_url: https://api.groq.com/openai/v1
+  api_key_env: GROQ_API_KEY
+
+openrouter:
+  base_url: https://openrouter.ai/api/v1
+  api_key_env: OPENROUTER_API_KEY
+
+# --- Self-hosted endpoints (no key by default) ---
+
+ollama:
+  base_url: http://localhost:11434/v1
+  api_key_env: OLLAMA_API_KEY
+  requires_key: false
+  notes: Local Ollama server.
+
+vllm:
+  base_url: http://localhost:8000/v1
+  api_key_env: VLLM_API_KEY
+  requires_key: false
+  notes: Self-hosted vLLM server speaking OpenAI's protocol.
+
+sglang:
+  base_url: http://localhost:30000/v1
+  api_key_env: SGLANG_API_KEY
+  requires_key: false
+  notes: Self-hosted SGLang server.
+
+lmstudio:
+  base_url: http://localhost:1234/v1
+  api_key_env: LMSTUDIO_API_KEY
+  requires_key: false
+  notes: LM Studio local server.
+
+local:
+  base_url: http://localhost:8000/v1
+  api_key_env: LOCAL_API_KEY
+  requires_key: false
+  notes: Catch-all entry for arbitrary local servers.
+
+# --- Add your own deployments below ---
+#
+# my-internal-vllm:
+#   base_url: https://vllm.internal.corp/v1
+#   api_key_env: INTERNAL_VLLM_KEY
+#   requires_key: true
+#   notes: Behind the corporate VPN.

graphgrail-0.1.0/configs/grail.yaml

@@ -0,0 +1,24 @@
+# GRAIL master config.
+# Loaded by `grail.config.load_config(path)`. Sibling per-module YAMLs in this
+# directory are auto-merged into the matching section, so you can split overrides
+# by concern. Documentation for every key lives in docs/<module>.md.
+
+project_name: example
+# Resolves with ~ expansion. All on-disk artefacts (parquets, lancedb, cache) live here.
+root_dir: ~/.grail/projects/example
+
+# Each block below mirrors a Pydantic submodel in grail/config.py. Anything you omit
+# falls back to the default in code. To split this file by module, leave the section
+# empty here and drop e.g. configs/llm.yaml next to this one — both load and merge.
+
+llm: {}
+embeddings: {}
+indexing: {}
+community: {}
+search: {}
+storage:
+  backend: local
+  # If unset, defaults to {root_dir}.
+  root: ~/.grail/projects/example
+prompts: {}
+vectorstore: {}

graphgrail-0.1.0/configs/indexing.yaml

@@ -0,0 +1,62 @@
+# Indexing pipeline config. See docs/indexing.md.
+
+# Chunking
+chunk_size: 2000
+chunk_overlap: 50
+encoding_name: cl100k_base
+document_boundary: "\n\n---DOCUMENT_BOUNDARY---\n\n"
+
+# Layout under storage root.
+input_folder: input
+output_folder: output
+cache_folder: cache
+
+# Per-stage endpoint+model overrides. When null, inherit from configs/llm.yaml.
+# Set one or both per stage — e.g. cheap extraction on Groq, headline summarization on OpenAI.
+entity_relation_endpoint: null
+entity_relation_model: null
+summarization_endpoint: null
+summarization_model: null
+
+# Entity types normalized to UPPER_SNAKE_CASE. Minimum 5 required.
+# Always include PERSON and ORGANIZATION as baseline types.
+# Add domain-specific types, or use discover_entity_types to let the LLM propose them.
+entity_types:
+  - person
+  - organization
+  - location
+  - event
+  - concept
+
+# When true, the LLM reads a corpus sample before extraction and proposes
+# additional entity types (merged with the list above, capped at max_entity_types).
+# Also available as CLI flag: grail index --discover-entities
+# discover_entity_types: false
+# max_entity_types: 15
+
+# Extraction delimiters — the entity/relationship parser is contract-bound to
+# these tokens. Change them only if you have a custom entity_relation prompt
+# that uses different delimiters. When absent, the built-in defaults apply.
+# tuple_delimiter: "<|>"
+# record_delimiter: "##"
+# completion_delimiter: "<|COMPLETE|>"
+# start_delimiter: "<|START_OUTPUT|>"
+
+# --- Token budgets per LLM call ---
+# extraction_max_tokens: max output tokens for each entity/relationship extraction call.
+# entity_discovery_max_tokens: max output tokens for the create-entities discovery call.
+# max_summarization_tokens: max tokens for entity description summarization.
+extraction_max_tokens: 8192
+entity_discovery_max_tokens: 2048
+max_summarization_tokens: 756
+
+# --- Per-stage concurrency ---
+# Override the global llm.concurrent_requests for specific stages.
+# null = use the global limit. Useful to run more extraction calls in parallel
+# (cheap, fast) while throttling heavier community-report generation.
+# extraction_concurrency: null
+# summarization_concurrency: null
+
+# Gleanings: the upstream graphrag setting that re-asks the LLM to find missed
+# entities. 0 = disabled. Costs roughly N× more tokens per chunk.
+max_gleanings: 0

graphgrail-0.1.0/configs/llm.yaml

@@ -0,0 +1,24 @@
+# LLM client config. See docs/llm.md.
+#
+# Endpoint and model are separate. `endpoint` is a name referenced in
+# configs/endpoints.yaml (defaults cover openai, anthropic, deepinfra, together,
+# groq, openrouter, ollama, vllm, sglang, lmstudio, local).
+
+# Which deployment of the OpenAI protocol to use by default.
+endpoint: openai
+# Model name within that endpoint. Whatever string the endpoint accepts.
+model: gpt-4o-mini
+
+# Concurrency / timeouts
+concurrent_requests: 15
+request_timeout: 180.0
+max_retries: 10
+max_retry_wait: 10.0
+sleep_on_rate_limit: 30.0
+
+debug: false
+
+# Disk cache. When enabled, identical (endpoint, model, messages, params) tuples
+# are reused. Useful during iteration.
+cache_enabled: false
+cache_dir: null  # null → {root_dir}/cache/llm

graphgrail-0.1.0/configs/prompts.yaml

@@ -0,0 +1,10 @@
+# Prompts config. See docs/prompts.md.
+
+# Extra directories searched (in order) before grail/prompts/builtin/. Each file
+# in these directories must expose NAME, REQUIRED_PARAMS, build_messages(**params).
+custom_paths: []
+
+# strict=true requires that the custom directory provide every built-in prompt
+# (entity_relation, summarize_description, community_report, json_correction,
+#  create_custom_entities, local_search, global_map, global_reduce, claim_extraction).
+strict: false

graphgrail-0.1.0/configs/reranker.yaml

@@ -0,0 +1,33 @@
+# Reranker config (optional). See docs/reranker.md.
+#
+# When enabled, a cross-encoder re-ranker scores entity and text-unit candidates
+# after the initial vector similarity retrieval. This improves which items fill
+# the limited context window sent to the LLM, at the cost of one extra API call.
+#
+# Toggle per-query with --rerank / --no-rerank on the CLI, or use_reranker= in
+# the Python API, without changing this config.
+
+enabled: false
+
+# Endpoint name (from endpoints.yaml) that hosts the reranker model.
+# The reranker uses a dedicated HTTP API (not OpenAI Chat Completions).
+endpoint: deepinfra
+
+# Cross-encoder model name. Passed to the inference API.
+model: Qwen/Qwen3-Reranker-0.6B
+
+# Full URL override. When null, auto-derived from endpoint name.
+# DeepInfra: https://api.deepinfra.com/v1/inference/{model}
+base_url: null
+
+# Over-fetch factor: retrieve top_k * overfetch_factor entities by vector
+# similarity, then rerank and trim to top_k. Higher = more candidates for the
+# reranker to choose from, but more tokens sent to the reranker API.
+overfetch_factor: 3
+
+# Which retrieval stages to rerank.
+rerank_entities: true
+rerank_text_units: true
+
+# HTTP timeout for the reranker API call (seconds).
+request_timeout: 30.0

graphgrail-0.1.0/configs/search.yaml

@@ -0,0 +1,35 @@
+# Search config. See docs/query.md.
+
+# Per-stage endpoint+model overrides. When null, inherit from configs/llm.yaml.
+local_search_endpoint: null
+local_search_model: null
+global_search_endpoint: null
+global_search_model: null
+
+# --- Local search budgeting ---
+local_max_tokens: 12000
+local_text_unit_prop: 0.5
+local_community_prop: 0.1
+local_conversation_history_max_turns: 5
+local_top_k_entities: 10
+local_top_k_relationships: 10
+
+# When false (default), community reports include full LLM-generated content
+# with detailed findings. When true, only the one-line summary is used (legacy
+# behaviour for tight token budgets or small-context models).
+use_community_summary: false
+
+# --- Global search budgeting ---
+global_map_max_tokens: 2000
+global_reduce_max_tokens: 8192
+# When the assembled community-report context exceeds this token budget, GRAIL
+# falls back to a map-reduce path. Otherwise it does a single reduce pass.
+global_chunk_size: 100000
+global_concurrency: 5
+
+# Maximum tokens the LLM can generate in its response.
+response_max_tokens: 16384
+
+# Description of the desired response style — injected into the global reduce
+# prompt's artifact_instructions block.
+response_type: "Multiple Paragraphs"

graphgrail-0.1.0/configs/storage.yaml

@@ -0,0 +1,14 @@
+# Storage backend config. See docs/storage.md.
+
+# "local" (default) writes everything under `root` on the filesystem.
+# "s3"    writes objects under `s3_bucket/s3_prefix` (requires the [s3] extra).
+backend: local
+
+# Local backend
+root: ~/.grail/projects/default
+
+# S3 backend (only used when backend=s3)
+s3_bucket: null
+s3_prefix: null
+s3_region: null
+s3_endpoint_url: null  # set for MinIO / S3-compatible stores

graphgrail-0.1.0/configs/templates/README.md

@@ -0,0 +1,101 @@
+# GRAIL configuration templates
+
+A **template** is a directory of YAML files matching the GRAIL config layout
+(``grail.yaml`` + optional ``endpoints.yaml``, ``llm.yaml``, ``embeddings.yaml``,
+``indexing.yaml``, ``community.yaml``, ``search.yaml``, ``storage.yaml``,
+``prompts.yaml``, ``vectorstore.yaml``).
+
+When you scaffold a project with ``grail init <project_dir> --template NAME``,
+every recognised YAML in that template directory is copied into the new project
+with ``{name}``, ``{root}``, and ``{date}`` placeholders substituted.
+
+## The "everything is OpenAI API" position
+
+GRAIL only speaks the OpenAI Chat Completions + Embeddings protocol. An
+**endpoint** is just one deployment of that protocol — a base URL the SDK posts
+to. Any compatible host plugs in: OpenAI itself, DeepInfra, Together, Groq,
+your own vLLM / SGLang / LM Studio server.
+
+The base URL is the ``/v1`` prefix. The SDK appends ``/chat/completions`` or
+``/embeddings`` automatically. **Do not include ``/chat/completions`` or
+``/embeddings`` in ``base_url``.**
+
+Examples of valid ``base_url`` values:
+
+| Host                  | base_url                                      |
+|-----------------------|------------------------------------------------|
+| OpenAI                | ``https://api.openai.com/v1``                  |
+| DeepInfra             | ``https://api.deepinfra.com/v1/openai``        |
+| Together              | ``https://api.together.xyz/v1``                |
+| Groq                  | ``https://api.groq.com/openai/v1``             |
+| Local vLLM            | ``http://localhost:8000/v1``                   |
+| Local SGLang          | ``http://localhost:30000/v1``                  |
+
+## Built-in templates
+
+### `low_cost_setup` — the one we ship
+
+A fully-filled GRAIL project biased for cost:
+
+- One inference endpoint pointed at DeepInfra (swap the ``base_url`` to retarget).
+- Conservative concurrency (8 in-flight calls).
+- LLM disk cache **on** so re-running indexing is free.
+- Single-pass extraction (``max_gleanings: 0``).
+- Tighter search token budgets than the GRAIL defaults.
+
+Files included:
+
+```
+configs/templates/low_cost_setup/
+├── grail.yaml             # project identity (placeholders for {name}/{root})
+├── endpoints.yaml         # one inference endpoint, shows the base_url pattern
+├── llm.yaml               # chat client + cache + concurrency
+├── embeddings.yaml        # embedding client (reuses the same endpoint)
+├── indexing.yaml          # chunking + extraction
+├── community.yaml         # Leiden + incremental thresholds + report generation
+├── search.yaml            # local + global search budgets
+├── storage.yaml           # local backend with {root} placeholder
+├── prompts.yaml           # custom prompt-pack search paths
+└── vectorstore.yaml       # LanceDB
+```
+
+Every key is documented in [docs/glossary.md](../../docs/glossary.md).
+
+## Using a template
+
+```bash
+grail init my-project --template low_cost_setup
+```
+
+After scaffolding, fill in your API key in `.env` (or the project's `.env`),
+edit `endpoints.yaml` if you want a different host, and you're ready to index.
+
+## Writing your own template
+
+1. Make a directory anywhere on disk, e.g. ``~/.grail/my-templates/acme/``.
+2. Copy any subset of the YAMLs above into it. ``grail.yaml`` is the minimum.
+3. Use placeholders ``{name}``, ``{root}``, ``{date}`` where you want them
+   substituted at init time. Other ``{…}`` sequences are passed through.
+4. Anything you don't include falls back to the defaults compiled into
+   ``grail/config.py``.
+
+Use with:
+
+```bash
+grail init my-project --template acme --templates-dir ~/.grail/my-templates
+```
+
+## Listing what's available
+
+```bash
+grail init --list-templates                              # built-ins only
+grail init --list-templates --templates-dir ~/.grail/    # plus your own
+```
+
+## Placeholders
+
+| Placeholder | Replaced with                                                |
+|-------------|---------------------------------------------------------------|
+| ``{name}``  | The ``--name`` flag (or project directory basename).          |
+| ``{root}``  | The absolute path of the new project directory.               |
+| ``{date}``  | ISO date at init time, e.g. ``2026-05-19``.                   |

graphgrail-0.1.0/configs/templates/low_cost_setup/community.yaml

@@ -0,0 +1,45 @@
+# Community detection (Leiden) + report generation.
+# See docs/incremental_pipeline.md for the algorithm and docs/glossary.md for keys.
+
+# --- Hierarchical Leiden ---
+# graspologic implementation. Larger max_cluster_size = fewer, broader communities.
+max_cluster_size: 50
+use_lcc: false              # don't restrict to the largest connected component
+strategy: leiden
+seed: 42                    # reproducible clustering — drop or change for variation
+
+# --- Small-community merging (DBSCAN over centroids) ---
+# Communities below min_community_size get merged with their nearest centroid.
+min_community_size: 10
+embedding_merge_eps: 0.5
+
+# --- Incremental updates ---
+# If (changed entities) / (total entities) >= threshold, re-cluster the affected
+# subgraph with Leiden. Below threshold, label-propagate new nodes onto existing
+# communities. 0.3 is the GraphRAG default; lower → more aggressive recompute.
+incremental_change_threshold: 0.3
+
+# --- Report generation (null = inherit from llm.yaml) ---
+community_report_endpoint: null
+community_report_model: null
+json_corrector_endpoint: null
+json_corrector_model: null
+
+# Output token cap for community reports.
+max_report_length: 4000
+
+# Concurrency for report LLM calls. null = use global llm.concurrent_requests.
+# report_concurrency: null
+
+# Claims (covariates) are off by default in v0.1 — pipeline not wired yet.
+include_covariates: false
+
+# --- Report-level filtering ---
+# Which Leiden hierarchy level to summarise. "coarsest" → smallest level number
+# (fewest, broadest communities; default). Other values: "finest", "all", or an int.
+community_level: coarsest
+
+# Skip communities with fewer than this many entities at report-generation time.
+# Defends against the long tail of singleton "communities" from isolated entities.
+# 0 disables the filter.
+min_report_size: 3

graphgrail-0.1.0/configs/templates/low_cost_setup/embeddings.yaml

@@ -0,0 +1,29 @@
+# Embedding client.
+#
+# Most providers serve chat AND embeddings under the same base_url — the OpenAI
+# SDK POSTs to {base_url}/embeddings automatically. So reusing the `inference`
+# endpoint here just works.
+#
+# If you ever need a different host for embeddings (separate server, different
+# region, …) define a second endpoint in endpoints.yaml and reference it below.
+
+endpoint: inference
+
+# The embedding model name your endpoint accepts. Qwen3-Embedding-0.6B is a
+# small, strong multilingual encoder available on DeepInfra.
+#
+# INVARIANT: this model must match between indexing and querying. Mismatched
+# dimensions silently break recall.
+model: Qwen/Qwen3-Embedding-0.6B
+
+encoding_format: float
+
+# Batching — most providers cap individual requests at a few hundred inputs.
+# 256 is safe across DeepInfra / OpenAI / Together. Raise if your provider allows.
+max_batch_size: 256
+concurrent_requests: 8
+
+request_timeout: 180.0
+max_retries: 10
+max_retry_wait: 10.0
+sleep_on_rate_limit: 30.0