uchi-python 0.1.0__tar.gz → 0.2.0__tar.gz

uchi_python-0.2.0/PKG-INFO

@@ -0,0 +1,273 @@
+Metadata-Version: 2.4
+Name: uchi-python
+Version: 0.2.0
+Summary: Omni-modal Deterministic Universal Sequence Predictor (ODUSP)
+License: MIT
+Project-URL: Documentation, https://github.com/JosephWoodall/uchi/tree/main/docs
+Project-URL: Homepage, https://github.com/JosephWoodall/uchi
+Keywords: machine learning,online learning,sequence prediction,time series,concept drift,trie,CTW
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT 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
+Classifier: Topic :: Software Development :: Libraries :: Python Modules
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: fastapi>=0.100.0
+Requires-Dist: uvicorn>=0.23.0
+Requires-Dist: requests>=2.31.0
+Requires-Dist: beautifulsoup4>=4.12.0
+Requires-Dist: tqdm>=4.66.0
+Requires-Dist: textual>=0.50.0
+Requires-Dist: torch>=2.0.0
+Requires-Dist: spacy>=3.0.0
+Provides-Extra: test
+Requires-Dist: pytest; extra == "test"
+Requires-Dist: pytest-cov; extra == "test"
+Requires-Dist: psutil; extra == "test"
+Provides-Extra: dev
+Requires-Dist: pytest; extra == "dev"
+Requires-Dist: pytest-cov; extra == "dev"
+Requires-Dist: ruff; extra == "dev"
+Requires-Dist: mypy; extra == "dev"
+Requires-Dist: pre-commit; extra == "dev"
+Requires-Dist: mkdocs-material; extra == "dev"
+Requires-Dist: nltk; extra == "dev"
+Requires-Dist: psutil; extra == "dev"
+Requires-Dist: datasets>=2.14.0; extra == "dev"
+Provides-Extra: sklearn
+Requires-Dist: scikit-learn>=1.0; extra == "sklearn"
+Provides-Extra: numpy
+Requires-Dist: numpy>=1.20; extra == "numpy"
+Provides-Extra: pandas
+Requires-Dist: pandas>=1.3; extra == "pandas"
+Provides-Extra: pyspark
+Requires-Dist: pyspark>=3.0.0; extra == "pyspark"
+Provides-Extra: optuna
+Requires-Dist: optuna>=3.0.0; extra == "optuna"
+Provides-Extra: faiss
+Requires-Dist: faiss-cpu>=1.7; extra == "faiss"
+Provides-Extra: all
+Requires-Dist: scikit-learn>=1.0; extra == "all"
+Requires-Dist: numpy>=1.20; extra == "all"
+Requires-Dist: pandas>=1.3; extra == "all"
+Requires-Dist: pyspark>=3.0.0; extra == "all"
+Requires-Dist: optuna>=3.0.0; extra == "all"
+Requires-Dist: pytest; extra == "all"
+Requires-Dist: faiss-cpu>=1.7; extra == "all"
+Dynamic: license-file
+
+![Uchi Logo](docs/logo.png)
+
+# Universal Sequence Predictor
+
+[![PyPI version](https://img.shields.io/pypi/v/uchi_python.svg)](https://pypi.org/project/uchi_python/)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
+[![Python Versions](https://img.shields.io/pypi/pyversions/uchi_python.svg)](https://pypi.org/project/uchi_python/)
+[![Tests](https://github.com/JosephWoodall/uchi/actions/workflows/ci.yml/badge.svg)](https://github.com/JosephWoodall/uchi/actions/workflows/ci.yml)
+
+## Core Mission: Omni-modal Deterministic Universal Sequence Predictor (ODUSP)
+Uchi v0.2.0 transforms the architecture from a simple sequence predictor into a completely multi-modal Deterministic Universal Sequence Predictor. It ingests text, audio, images, math telemetry, and code simultaneously—without any neural weights or pre-training. It adds a structured routing layer with intent-based query dispatch via `ProceduralMemory`, a trainable SSM confidence signal via GRPO, persistent vector memory, TUI, REST API, and SDK — all without introducing an LLM dependency.
+
+
+> [!NOTE]
+> Please see `docs/` for the complete Algorithmic Walkthrough, ODUSP vs LLM Benchmarks, and full API references.
+
+---
+
+> [!NOTE]
+> **Comprehensive Documentation & API Reference**
+>
+> For interactive examples, API documentation, and to see the newest capabilities (including online Math Learning, Vector Retrievals, and the Simulation Engine), please see our full documentation website.
+>
+> **[Read the Full Documentation →](https://github.com/JosephWoodall/uchi/tree/main/docs)**
+
+---
+
+## Installation
+
+```bash
+pip install -e ".[all]"
+```
+
+On first launch, Uchi runs a one-time bootstrap (Python stdlib patterns + Wikipedia facts) and saves the result to `brain.uchi`. Subsequent launches are instant.
+
+## Quickstart
+
+Uchi has two entry points that share the same `brain.uchi` — every interaction in either one improves the other.
+
+### 1. Terminal UI (TUI)
+
+```bash
+uchi                        # launch interactive chat
+uchi --preload data.txt     # pre-train with a file before chatting
+uchi --brain /path/to/brain.uchi   # use a specific brain file
+```
+
+Inside the TUI:
+
+| Command | Description |
+|---|---|
+| Just type | Chat with Uchi — it learns from every turn |
+| `/load <file>` | Stream any file into the knowledge base |
+| `/save` | Force-save the current brain state to disk |
+| `Ctrl+S` | Save brain |
+| `Ctrl+C` | Save and quit |
+
+Uchi gives positive/negative feedback signals to improve itself — type "good", "correct", "yes" to reinforce a response, or "wrong", "bad", "no" to prune it.
+
+### 2. REST API
+
+```bash
+uvicorn uchi.api_server:app --host 0.0.0.0 --port 8000
+```
+
+**POST /chat**
+```bash
+curl -X POST http://localhost:8000/chat \
+  -H "Content-Type: application/json" \
+  -d '{"message": "what is the capital of France?"}'
+# {"reply": "paris", "entropy": 3.2}
+```
+
+**GET /metrics**
+```bash
+curl http://localhost:8000/metrics
+# {"status": "online", "memory_records": 1024, "mode": "deterministic"}
+```
+
+**GET /debug/walk**
+```bash
+curl http://localhost:8000/debug/walk
+# Returns trie walk data from the last prediction (depth, contributions, similarity)
+```
+
+### 3. Python API
+
+```python
+from uchi.omni_router import OmniRouter
+from uchi.cli import load_brain, save_brain
+
+# Load existing brain or create a new one
+router = load_brain("brain.uchi") or OmniRouter()
+
+# Chat
+reply = router.chat("what is the capital of France?")
+print(reply)  # → "paris"
+
+# Teach it something new
+router.stream(["<|user|>", "what", "is", "the", "capital", "of", "germany",
+               "<|assistant|>", "berlin"])
+
+# Save
+save_brain(router, "brain.uchi")
+```
+
+### 4. Offline Knowledge Bootstrapping
+
+To scale Uchi's knowledge base beyond the cold-start defaults, run these scripts once before distributing your `brain.uchi`:
+
+```bash
+# Ingest Wikipedia + code_search_net via HuggingFace (requires: pip install datasets)
+python scripts/bootstrap_knowledge.py --limit 10000
+
+# Ingest Python stdlib function patterns via AST (no internet required)
+python scripts/bootstrap_code.py
+
+# Ingest Wikipedia fact triples via spaCy SVO extraction (requires: pip install wikipedia spacy)
+python scripts/bootstrap_wikidata.py
+```
+
+The resulting `brain.uchi` can be committed to your repo or distributed with your package so end users start with a pre-trained brain.
+
+## Benchmarks
+
+Uchi is a **deterministic sequence predictor**, not a language model. Its benchmarks measure properties that LLMs cannot demonstrate — not perplexity or few-shot accuracy, but whether a system that has *seen* a fact will *deterministically recall* it, resist overwriting it under noise, and stay fast as its knowledge base scales.
+
+Run yourself with:
+```bash
+python benchmarks/run_benchmarks.py
+python benchmarks/run_benchmarks.py --mini       # fast CI pass (10 facts)
+python benchmarks/run_benchmarks.py --wipe       # clean rebuild before benchmarking
+```
+
+Results are written to `eval_metrics.json` and this table is updated automatically.
+
+---
+
+### Pre-load Recall — **80.0%** (40 / 50)
+
+50 factual Q&A pairs (geography, science, history, Python/CS) are streamed directly into the trie as `<|user|> question <|assistant|> answer` sequences. Web search is then disabled and the system is asked each question cold. A pass requires the expected answer to appear in the reply.
+
+This is Uchi's core capability claim: *if you teach it something, it recalls it exactly*. The 80% figure reflects the current pipeline correctness across a diverse fact set including multi-word answers, numeric values, and chemical symbols. Failures are vocabulary edge cases (the tokenizer normalises "au" → `gold.n.03`, which is semantically correct but fails substring match).
+
+---
+
+### Zero Catastrophic Forgetting — **100.0%** (10 / 10, after 1 000 noise facts)
+
+10 anchor facts are streamed first. Then 1 000 unrelated noise facts are streamed on top. The 10 anchors are then re-tested. 100% means not a single anchor fact was displaced.
+
+LLMs trained on a new document lose previously learned facts proportional to the dataset shift (catastrophic interference). Uchi uses a prefix trie: new paths are inserted without touching existing ones. Recall of any fact streamed in the past is bounded only by trie depth, not by how much has been streamed since.
+
+---
+
+### Latency vs. Brain Size — flat O(depth)
+
+| Brain size | Latency |
+|---|---|
+| 10 facts | 10 666 ms |
+| 100 facts | 2 568 ms |
+| 500 facts | 2 282 ms |
+| 1 000 facts | 2 597 ms |
+
+Latency is measured as wall-clock time for a single chat() turn on a pre-loaded fact, with web search disabled.
+
+The pattern is deliberate: latency at 1 000 facts is the same as at 100 facts because trie lookup is O(depth), not O(vocabulary size). The 10-fact spike reflects cold-start overhead (first MCTS warmup before the loop has converged). At scale this overhead amortises to near-zero.
+
+---
+
+### Code Completion — **5.0%** (1 / 20 HumanEval)
+
+20 HumanEval function stubs (`def factorial(n):` etc.) are streamed as training pairs, then recalled. Scored by `TieredCodeOracle`: the generated body must parse as valid Python (`ast.parse`) and contain expected keywords.
+
+5% on HumanEval after single-pass training is the *floor*, not the ceiling. Uchi is not pre-trained on code corpora. The 1/20 passing case demonstrates that the code recall pipeline is functional end-to-end. Higher scores require either multiple training passes or the `brain_code.uchi` specialist loaded alongside the base brain.
+
+---
+
+### Inference Latency — **2 333 ms** per turn
+
+Single chat turn on a pre-loaded fact, web search off. This exercises the full pipeline: tokenise → trie peek → pre-flight classify → greedy bypass → CoherenceOracle → detokenise. Down from 17 762 ms in the pre-optimization baseline (7.6× faster) after dynamic MCTS budget scaling: factual queries now exit via O(1) greedy bypass instead of running the full 20-rollout MCTS loop.
+
+---
+
+### RAM Footprint — **1 374 MB** resident
+
+Measured after loading `brain.uchi` and running the recall stream. Dominated by the trie node store (~1.1 GB for the pre-built brain) plus the SSM embedding table (~180 MB at d_model=256). The trie is the canonical in-memory database; no separate vector store is required for retrieval.
+
+---
+
+### Hallucination Rate — **0%**
+
+Uchi cannot fabricate tokens that are not in its trie. Every generated token is drawn from the empirical distribution at a trie node that was built from real streamed data. The CoherenceOracle enforces a secondary check (overlap, trigram repetition, SSM gate) and returns `[Uncertain]` rather than confabulate when no valid candidate passes. Zero hallucination is a structural guarantee, not a tuned behaviour.
+
+---
+
+<!-- BENCHMARK_TABLE_START -->
+| Metric | Score | Notes |
+|---|---|---|
+| **Pre-load Recall** | **80.0%** (n=50) | Stream N facts → immediately test recall; measures deterministic memory |
+| **Zero Catastrophic Forgetting** | **100.0%** after 1000 noise facts | Anchor facts recalled correctly after 1000 distractors streamed on top |
+| **Latency vs. Brain Size** | 10facts→10666ms  100facts→2568ms  500facts→2282ms  1000facts→2597ms | Proves O(depth) trie lookup: latency stays flat as brain grows |
+| **Code Completion** | **5.0%** (n=20 HumanEval) | Python function stub → body; scored by syntax + keyword validity |
+| **Inference Latency** | **2333.1 ms** | Single turn on a pre-loaded fact, web search disabled |
+| **RAM Footprint** | **1374.2 MB** | Resident set after brain load + recall stream |
+| **Hallucination Rate** | **0%** | Strict trie boundary enforcement |
+<!-- BENCHMARK_TABLE_END -->
+
+
+

uchi_python-0.2.0/README.md

@@ -0,0 +1,209 @@
+![Uchi Logo](docs/logo.png)
+
+# Universal Sequence Predictor
+
+[![PyPI version](https://img.shields.io/pypi/v/uchi_python.svg)](https://pypi.org/project/uchi_python/)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
+[![Python Versions](https://img.shields.io/pypi/pyversions/uchi_python.svg)](https://pypi.org/project/uchi_python/)
+[![Tests](https://github.com/JosephWoodall/uchi/actions/workflows/ci.yml/badge.svg)](https://github.com/JosephWoodall/uchi/actions/workflows/ci.yml)
+
+## Core Mission: Omni-modal Deterministic Universal Sequence Predictor (ODUSP)
+Uchi v0.2.0 transforms the architecture from a simple sequence predictor into a completely multi-modal Deterministic Universal Sequence Predictor. It ingests text, audio, images, math telemetry, and code simultaneously—without any neural weights or pre-training. It adds a structured routing layer with intent-based query dispatch via `ProceduralMemory`, a trainable SSM confidence signal via GRPO, persistent vector memory, TUI, REST API, and SDK — all without introducing an LLM dependency.
+
+
+> [!NOTE]
+> Please see `docs/` for the complete Algorithmic Walkthrough, ODUSP vs LLM Benchmarks, and full API references.
+
+---
+
+> [!NOTE]
+> **Comprehensive Documentation & API Reference**
+>
+> For interactive examples, API documentation, and to see the newest capabilities (including online Math Learning, Vector Retrievals, and the Simulation Engine), please see our full documentation website.
+>
+> **[Read the Full Documentation →](https://github.com/JosephWoodall/uchi/tree/main/docs)**
+
+---
+
+## Installation
+
+```bash
+pip install -e ".[all]"
+```
+
+On first launch, Uchi runs a one-time bootstrap (Python stdlib patterns + Wikipedia facts) and saves the result to `brain.uchi`. Subsequent launches are instant.
+
+## Quickstart
+
+Uchi has two entry points that share the same `brain.uchi` — every interaction in either one improves the other.
+
+### 1. Terminal UI (TUI)
+
+```bash
+uchi                        # launch interactive chat
+uchi --preload data.txt     # pre-train with a file before chatting
+uchi --brain /path/to/brain.uchi   # use a specific brain file
+```
+
+Inside the TUI:
+
+| Command | Description |
+|---|---|
+| Just type | Chat with Uchi — it learns from every turn |
+| `/load <file>` | Stream any file into the knowledge base |
+| `/save` | Force-save the current brain state to disk |
+| `Ctrl+S` | Save brain |
+| `Ctrl+C` | Save and quit |
+
+Uchi gives positive/negative feedback signals to improve itself — type "good", "correct", "yes" to reinforce a response, or "wrong", "bad", "no" to prune it.
+
+### 2. REST API
+
+```bash
+uvicorn uchi.api_server:app --host 0.0.0.0 --port 8000
+```
+
+**POST /chat**
+```bash
+curl -X POST http://localhost:8000/chat \
+  -H "Content-Type: application/json" \
+  -d '{"message": "what is the capital of France?"}'
+# {"reply": "paris", "entropy": 3.2}
+```
+
+**GET /metrics**
+```bash
+curl http://localhost:8000/metrics
+# {"status": "online", "memory_records": 1024, "mode": "deterministic"}
+```
+
+**GET /debug/walk**
+```bash
+curl http://localhost:8000/debug/walk
+# Returns trie walk data from the last prediction (depth, contributions, similarity)
+```
+
+### 3. Python API
+
+```python
+from uchi.omni_router import OmniRouter
+from uchi.cli import load_brain, save_brain
+
+# Load existing brain or create a new one
+router = load_brain("brain.uchi") or OmniRouter()
+
+# Chat
+reply = router.chat("what is the capital of France?")
+print(reply)  # → "paris"
+
+# Teach it something new
+router.stream(["<|user|>", "what", "is", "the", "capital", "of", "germany",
+               "<|assistant|>", "berlin"])
+
+# Save
+save_brain(router, "brain.uchi")
+```
+
+### 4. Offline Knowledge Bootstrapping
+
+To scale Uchi's knowledge base beyond the cold-start defaults, run these scripts once before distributing your `brain.uchi`:
+
+```bash
+# Ingest Wikipedia + code_search_net via HuggingFace (requires: pip install datasets)
+python scripts/bootstrap_knowledge.py --limit 10000
+
+# Ingest Python stdlib function patterns via AST (no internet required)
+python scripts/bootstrap_code.py
+
+# Ingest Wikipedia fact triples via spaCy SVO extraction (requires: pip install wikipedia spacy)
+python scripts/bootstrap_wikidata.py
+```
+
+The resulting `brain.uchi` can be committed to your repo or distributed with your package so end users start with a pre-trained brain.
+
+## Benchmarks
+
+Uchi is a **deterministic sequence predictor**, not a language model. Its benchmarks measure properties that LLMs cannot demonstrate — not perplexity or few-shot accuracy, but whether a system that has *seen* a fact will *deterministically recall* it, resist overwriting it under noise, and stay fast as its knowledge base scales.
+
+Run yourself with:
+```bash
+python benchmarks/run_benchmarks.py
+python benchmarks/run_benchmarks.py --mini       # fast CI pass (10 facts)
+python benchmarks/run_benchmarks.py --wipe       # clean rebuild before benchmarking
+```
+
+Results are written to `eval_metrics.json` and this table is updated automatically.
+
+---
+
+### Pre-load Recall — **80.0%** (40 / 50)
+
+50 factual Q&A pairs (geography, science, history, Python/CS) are streamed directly into the trie as `<|user|> question <|assistant|> answer` sequences. Web search is then disabled and the system is asked each question cold. A pass requires the expected answer to appear in the reply.
+
+This is Uchi's core capability claim: *if you teach it something, it recalls it exactly*. The 80% figure reflects the current pipeline correctness across a diverse fact set including multi-word answers, numeric values, and chemical symbols. Failures are vocabulary edge cases (the tokenizer normalises "au" → `gold.n.03`, which is semantically correct but fails substring match).
+
+---
+
+### Zero Catastrophic Forgetting — **100.0%** (10 / 10, after 1 000 noise facts)
+
+10 anchor facts are streamed first. Then 1 000 unrelated noise facts are streamed on top. The 10 anchors are then re-tested. 100% means not a single anchor fact was displaced.
+
+LLMs trained on a new document lose previously learned facts proportional to the dataset shift (catastrophic interference). Uchi uses a prefix trie: new paths are inserted without touching existing ones. Recall of any fact streamed in the past is bounded only by trie depth, not by how much has been streamed since.
+
+---
+
+### Latency vs. Brain Size — flat O(depth)
+
+| Brain size | Latency |
+|---|---|
+| 10 facts | 10 666 ms |
+| 100 facts | 2 568 ms |
+| 500 facts | 2 282 ms |
+| 1 000 facts | 2 597 ms |
+
+Latency is measured as wall-clock time for a single chat() turn on a pre-loaded fact, with web search disabled.
+
+The pattern is deliberate: latency at 1 000 facts is the same as at 100 facts because trie lookup is O(depth), not O(vocabulary size). The 10-fact spike reflects cold-start overhead (first MCTS warmup before the loop has converged). At scale this overhead amortises to near-zero.
+
+---
+
+### Code Completion — **5.0%** (1 / 20 HumanEval)
+
+20 HumanEval function stubs (`def factorial(n):` etc.) are streamed as training pairs, then recalled. Scored by `TieredCodeOracle`: the generated body must parse as valid Python (`ast.parse`) and contain expected keywords.
+
+5% on HumanEval after single-pass training is the *floor*, not the ceiling. Uchi is not pre-trained on code corpora. The 1/20 passing case demonstrates that the code recall pipeline is functional end-to-end. Higher scores require either multiple training passes or the `brain_code.uchi` specialist loaded alongside the base brain.
+
+---
+
+### Inference Latency — **2 333 ms** per turn
+
+Single chat turn on a pre-loaded fact, web search off. This exercises the full pipeline: tokenise → trie peek → pre-flight classify → greedy bypass → CoherenceOracle → detokenise. Down from 17 762 ms in the pre-optimization baseline (7.6× faster) after dynamic MCTS budget scaling: factual queries now exit via O(1) greedy bypass instead of running the full 20-rollout MCTS loop.
+
+---
+
+### RAM Footprint — **1 374 MB** resident
+
+Measured after loading `brain.uchi` and running the recall stream. Dominated by the trie node store (~1.1 GB for the pre-built brain) plus the SSM embedding table (~180 MB at d_model=256). The trie is the canonical in-memory database; no separate vector store is required for retrieval.
+
+---
+
+### Hallucination Rate — **0%**
+
+Uchi cannot fabricate tokens that are not in its trie. Every generated token is drawn from the empirical distribution at a trie node that was built from real streamed data. The CoherenceOracle enforces a secondary check (overlap, trigram repetition, SSM gate) and returns `[Uncertain]` rather than confabulate when no valid candidate passes. Zero hallucination is a structural guarantee, not a tuned behaviour.
+
+---
+
+<!-- BENCHMARK_TABLE_START -->
+| Metric | Score | Notes |
+|---|---|---|
+| **Pre-load Recall** | **80.0%** (n=50) | Stream N facts → immediately test recall; measures deterministic memory |
+| **Zero Catastrophic Forgetting** | **100.0%** after 1000 noise facts | Anchor facts recalled correctly after 1000 distractors streamed on top |
+| **Latency vs. Brain Size** | 10facts→10666ms  100facts→2568ms  500facts→2282ms  1000facts→2597ms | Proves O(depth) trie lookup: latency stays flat as brain grows |
+| **Code Completion** | **5.0%** (n=20 HumanEval) | Python function stub → body; scored by syntax + keyword validity |
+| **Inference Latency** | **2333.1 ms** | Single turn on a pre-loaded fact, web search disabled |
+| **RAM Footprint** | **1374.2 MB** | Resident set after brain load + recall stream |
+| **Hallucination Rate** | **0%** | Strict trie boundary enforcement |
+<!-- BENCHMARK_TABLE_END -->
+
+
+

uchi_python-0.2.0/pyproject.toml

@@ -0,0 +1,86 @@
+[build-system]
+requires = ["setuptools>=61.0.0", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "uchi-python"
+version = "0.2.0"
+description = "Omni-modal Deterministic Universal Sequence Predictor (ODUSP)"
+readme = "README.md"
+requires-python = ">=3.10"
+license = { text = "MIT" }
+keywords = [
+    "machine learning", "online learning", "sequence prediction",
+    "time series", "concept drift", "trie", "CTW",
+]
+
+classifiers = [
+    "Development Status :: 3 - Alpha",
+    "Intended Audience :: Science/Research",
+    "License :: OSI Approved :: MIT License",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+    "Topic :: Software Development :: Libraries :: Python Modules",
+]
+
+# Pure Python — no required dependencies.
+# Pure Python core dependencies
+dependencies = [
+    "fastapi>=0.100.0",
+    "uvicorn>=0.23.0",
+    "requests>=2.31.0",
+    "beautifulsoup4>=4.12.0",
+    "tqdm>=4.66.0",
+    "textual>=0.50.0",
+    "torch>=2.0.0",
+    "spacy>=3.0.0",
+]
+
+[project.urls]
+Documentation = "https://github.com/JosephWoodall/uchi/tree/main/docs"
+Homepage = "https://github.com/JosephWoodall/uchi"
+
+[project.scripts]
+uchi = "uchi.cli:main"
+
+[project.optional-dependencies]
+test    = ["pytest", "pytest-cov", "psutil"]
+dev     = ["pytest", "pytest-cov", "ruff", "mypy", "pre-commit", "mkdocs-material", "nltk", "psutil", "datasets>=2.14.0"]
+sklearn = ["scikit-learn>=1.0"]
+numpy   = ["numpy>=1.20"]
+pandas  = ["pandas>=1.3"]
+pyspark = ["pyspark>=3.0.0"]
+optuna  = ["optuna>=3.0.0"]
+faiss   = ["faiss-cpu>=1.7"]
+all     = ["scikit-learn>=1.0", "numpy>=1.20", "pandas>=1.3", "pyspark>=3.0.0", "optuna>=3.0.0", "pytest", "faiss-cpu>=1.7"]
+
+[tool.setuptools.packages.find]
+where   = ["."]
+include = ["uchi*"]
+
+[tool.setuptools.package-data]
+uchi = ["py.typed", "persona.txt"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+markers = [
+    "eval: quality benchmark tests — slow, excluded from normal CI (run with: pytest -m eval)",
+]
+
+[tool.ruff]
+line-length = 100
+target-version = "py310"
+exclude = [".git", ".venv", "__pycache__"]
+
+[tool.mypy]
+python_version = "3.10"
+warn_return_any = false
+warn_unused_configs = true
+ignore_missing_imports = true
+
+[[tool.mypy.overrides]]
+module = ["numpy.*"]
+ignore_errors = true