| --- |
| license: apache-2.0 |
| tags: |
| - graph-reasoning |
| - non-neural |
| - cpu-native |
| - explainable-ai |
| - vector-symbolic |
| language: |
| - en |
| - multilingual |
| library_name: webmind |
| pipeline_tag: text-generation |
| --- |
| |
| # webmind-brain-v1 |
|
|
| A graph-based reasoning engine. Not a neural network. No gradient descent. No GPU required. |
|
|
| The brain learns by building a co-occurrence graph over word vectors, then reasons by converging through the graph. Every answer has a traceable source. Knowledge is editable and deletable. |
|
|
| ## Quick Start |
|
|
| ```bash |
| pip install numpy fastapi uvicorn lmdb |
| ``` |
|
|
| ```python |
| from webmind import Brain |
| |
| brain = Brain.from_pretrained("webmind/webmind-brain-v1") |
| |
| # Teach it something |
| brain.teach("Paris is the capital of France") |
| brain.teach("London is the capital of England") |
| |
| # Ask |
| result = brain.ask("capital of France") |
| print(result["answer"]) # paris capital france |
| print(result["confidence"]) # 0.85 |
| print(result["strategy"]) # convergence / co-occurrence / abstain |
| |
| # Generate fluent text |
| gen = brain.generate("Tell me about France", max_tokens=20, temperature=0.7) |
| print(gen["text"]) |
| |
| # Save |
| brain.flush() |
| ``` |
|
|
| ## OpenAI-Compatible Server |
|
|
| ```bash |
| python serve.py |
| # Then: |
| curl http://localhost:8000/v1/chat/completions \ |
| -H "Content-Type: application/json" \ |
| -d '{"messages": [{"role": "user", "content": "capital of france"}]}' |
| ``` |
|
|
| Supports streaming (`"stream": true`), the `/v1/models` endpoint, and `/health`. |
|
|
| ## Architecture |
|
|
| ``` |
| Input -> Garbage Filter (heuristic + LSH) |
| -> Tier 1: Q→A Direct Lookup (LRU + LMDB, <1ms) |
| -> Tier 1.5: LSH Semantic Search (O(1) bucket lookup, seed concepts) |
| -> Tier 2: Convergence Loop (multi-hop reasoning over sparse graph) |
| -> Co-occurrence Search (complementary sparse signal) |
| -> Sentence Retrieval (full text from LMDB) |
| -> Confidence Floor (abstain if < 0.15) |
| -> Web Search fallback (DuckDuckGo + Wikipedia) |
| ``` |
|
|
| Key properties: |
| - **Co-occurrence graph**: words that appear together pull toward each other in a sparse matrix |
| - **Convergence loop**: iteratively search the graph, blending discovered concepts back into the query until the output stabilizes |
| - **Dual retrieval**: dense neuron search + sparse co-occurrence search race in parallel |
| - **Successor chains**: each word neuron stores its top-10 successors for generation |
| - **Confidence tracking**: every neuron has a confidence score that grows when useful and shrinks when not |
| - **LSH vocabulary filter**: locality-sensitive hashing over MiniLM embeddings for garbage detection, morphological linking ("gravitational"→"gravity"), vocabulary dedup, and O(1) semantic search |
| - **ScaNN backend**: Google's anisotropic vector quantization for faster ANN search (optional, falls back to LSH) |
| - **Int8 quantization**: PolarQuant-inspired 4x embedding compression with ~1% accuracy loss |
| - **Confidence floor**: abstain rather than return weak convergence results (bad context > no context) |
| - **Vocabulary pruning**: score words by convergence contribution, remove low-value entries |
|
|
| ## What It Is Good At |
|
|
| - Factual Q&A with traceable sources |
| - Multi-hop reasoning (convergence crosses concept boundaries) |
| - Incremental learning (teach new facts at runtime, no retraining) |
| - Honest failure (says "I don't know" when it doesn't converge) |
| - Knowledge editing (delete a neuron = delete a fact) |
|
|
| ## What It Is Not Good At |
|
|
| - Fluent prose generation (output is concept-oriented, not grammatically polished) |
| - Creative writing |
| - Long-form text |
| - Tasks requiring deep syntactic understanding |
|
|
| ## Training Data |
|
|
| This model ships empty. It learns from what you teach it. The `from_pretrained` download includes the graph structure and vocabulary but no pre-loaded knowledge. |
|
|
| For evaluation, we tested on HotPotQA (200 train, 50 test) achieving 72% exact match with word neurons + successor chains. |
|
|
| ## Limitations |
|
|
| - Context window is limited by the convergence loop (not fixed-length, but practically ~10 hops) |
| - Generation quality depends heavily on what has been taught |
| - No coreference resolution beyond what convergence provides |
| - Function words are stripped during reasoning (grammar handled separately) |
|
|
| ## Citation |
|
|
| If you use this work, please cite: |
|
|
| ```bibtex |
| @software{webmind_brain_2026, |
| title={Webmind Brain: Graph-Based Reasoning Without Neural Networks}, |
| url={https://github.com/webmind-ai/webmind-brain}, |
| year={2026}, |
| license={Apache-2.0} |
| } |
| ``` |
|
|
| ## License |
|
|
| Apache 2.0 |
|
|
