agentinputlab 0.1.0a1__py3-none-any.whl

workers/reformer/README.md

@@ -0,0 +1,125 @@
+# ReFormeR pinned inference worker
+
+This worker executes only the separable inference path from
+[`aminbigdeli/ReFormeR`](https://github.com/aminbigdeli/ReFormeR/tree/72e52450a922bc0051b5b39c3a6307186a555137).
+It loads the official `Reformulator` class from a user-mounted checkout and
+directly calls `select_best_pattern()` and `apply_pattern()`. It does not vendor
+or copy upstream source, run pattern induction, or claim paper-result parity.
+
+## License gate
+
+The pinned repository has no root `LICENSE`, `COPYING`, or `NOTICE` file. Its
+README contains an MIT badge, but a badge is not a complete license grant, so
+AgentInputLab records the source license as `NOASSERTION`.
+
+Execution requires two explicit acknowledgements:
+
+1. the adapter sets `source_license_acknowledged=True` per request;
+2. the operator starts the worker with `--acknowledge-noassertion-license`.
+
+The worker verifies the exact commit, README evidence hash, every declared
+source hash, the continued absence of a license file, and a clean Git checkout.
+If the upstream project later adds a license, the current manifest deliberately
+fails closed until it is re-audited.
+
+## Four non-interchangeable profiles
+
+| Release | Pattern choice | Materialized query |
+|---|---|---|
+| `rewrite.reformer.repo_native_r_only` | official automatic selection | repository `r` only |
+| `rewrite.reformer.paper_spec_q_concat_r` | official automatic selection | paper `q ⊕ r` |
+| `rewrite.reformer.controlled_r_only` | manually fixed pattern, official application | `r` only |
+| `rewrite.reformer.controlled_q_concat_r` | manually fixed pattern, official application | `q ⊕ r` |
+
+Automatic selection calls the official selector and application methods.
+Manual profiles are a controlled ablation: AgentInputLab fixes a named pattern
+from the exact hashed pattern set, then calls only the official application
+method. The `q ⊕ r` profiles concatenate outside upstream code because the
+repository returns `r` while the paper specifies concatenation. These axes are
+recorded in the release ID, request identity, response, trace, and patch hash.
+
+## Bound comparison state
+
+Every request binds:
+
+- source commit and all relevant source hashes;
+- prompt JSON SHA-256 and pattern-file byte SHA-256;
+- the complete selected pattern, including its examples;
+- model deployment fingerprint and retrieval fingerprint;
+- ordered document IDs/text and their canonical retrieval-state hash;
+- manual pattern ID when applicable;
+- seed plus model-call, provider-token, and priced-cost ceilings.
+
+Endpoint, model, API-key environment variable, checkout, and pattern path are
+startup-only settings. JSONL requests cannot override them.
+
+## Start the worker
+
+Install the inference worker's source-import dependency in its isolated Python
+environment. vLLM and upstream training dependencies are not required by this
+remote-model seam:
+
+```bash
+pip install 'agentinputlab[reformer]'
+```
+
+First verify a clean checkout without importing vLLM or making a model call:
+
+```bash
+python workers/reformer/driver.py \
+  --checkout /opt/upstreams/ReFormeR \
+  --verify-only
+```
+
+Then serve it through an OpenAI-compatible endpoint. This endpoint may be a
+hosted API or a local vLLM server exposing the OpenAI protocol.
+
+```bash
+python workers/reformer/driver.py \
+  --checkout /opt/upstreams/ReFormeR \
+  --patterns /opt/artifacts/extracted_patterns.json \
+  --acknowledge-noassertion-license \
+  --backend-config-json "$REFORMER_BACKEND_CONFIG"
+```
+
+Example backend configuration:
+
+```json
+{
+  "base_url": "http://127.0.0.1:8000/v1",
+  "api_key_env": "REFORMER_API_KEY",
+  "model": "Qwen/Qwen2.5-7B-Instruct",
+  "fingerprint": {
+    "provider": "vllm",
+    "model": "Qwen/Qwen2.5-7B-Instruct",
+    "revision": "deployment-or-weight-revision"
+  },
+  "extra_body": {},
+  "timeout_seconds": 60,
+  "max_input_tokens_per_call": 28000,
+  "max_output_tokens": 500,
+  "temperature": 0.1,
+  "input_cost_per_million": 0,
+  "output_cost_per_million": 0
+}
+```
+
+The worker's `max_tokens` control covers provider **input plus output** tokens;
+the shared runner's `max_output_tokens` budget covers generated output. For the
+startup limits above, configure the automatic profile with
+`ReFormeRControls(max_model_calls=2, max_tokens=2 * (28000 + 500))`. Manual
+profiles need one such call allowance and an explicit `manual_pattern_id` from
+the hashed pattern library. Do not raise the shared output budget to account for
+input tokens or count unpriced local generation as zero compute cost.
+
+## Deliberate limits
+
+- Pattern induction symbols and hashes are audited, but induction is not
+  executed by this inference worker.
+- The official vLLM chat-template serialization is replaced by a transparent
+  OpenAI-compatible message seam; prompt content still comes from the pinned
+  official `PromptManager` and `prompts.json`.
+- The repository's invalid-selection and no-document first-pattern fallbacks
+  are preserved and surfaced explicitly.
+- No official output fixture or benchmark-result parity is available yet, so
+  all four profiles remain `external_only`.

workers/reformer/__init__.py

@@ -0,0 +1 @@
+"""Pinned user-mounted ReFormeR inference worker."""

workers/reformer/compatibility_manifest.json

@@ -0,0 +1,68 @@
+{
+  "schema_version": "agentinputlab.reformer.compatibility.v1",
+  "upstream": {
+    "repository": "https://github.com/aminbigdeli/ReFormeR.git",
+    "commit": "72e52450a922bc0051b5b39c3a6307186a555137",
+    "license_spdx": "NOASSERTION",
+    "license_file_present": false,
+    "license_evidence_path": "README.md",
+    "license_evidence_sha256": "f8030ec4fd3de9ef43fafe78cf51fc7f71316f60acdf621ff6b52606920dabef",
+    "require_clean_checkout": true
+  },
+  "source_files": {
+    "README.md": "f8030ec4fd3de9ef43fafe78cf51fc7f71316f60acdf621ff6b52606920dabef",
+    "requirements.txt": "6c3cd77a20101d96b729a8965e130ab41137e4107a9f6601edfc0555b8fedbfd",
+    "reformer/core/reformer.py": "1cacdc8dca84d863e4f559e2f078c777631e92f3b4cc7d6de6d3b76b31d592c5",
+    "reformer/prompt_manager.py": "7f154ca601470ce1918ebf2d72552fbef9d69bb21e78b6782fd94ac07e134d33",
+    "reformer/prompts.json": "536bd3bd9596baf65c6b8f98f05e1320d59f221ae9b8d1799461374cd3fad652",
+    "reformer/patterns/extracted_patterns.json": "e2d0b1c7edc73c82776a1f01fc03f0ec5b06f286ac178ba5577cc0865d9bcaa5",
+    "reformer/patterns/query_reformulation_prompts.py": "09a5e8e3c8757fb03c4deefb82a5544c3f53e3f346aa54b275dcf7f833ff3d1d",
+    "reformer/patterns/iterative_pattern_extraction.py": "c12709be03c5ae8b96332d00ad24495fdff9fc5121a177069627dbd6d7edef96"
+  },
+  "official_symbols": {
+    "runtime": {
+      "path": "reformer/core/reformer.py",
+      "class": "Reformulator",
+      "methods": [
+        "select_best_pattern",
+        "apply_pattern",
+        "reformulate_query"
+      ]
+    },
+    "induction_audit_only": {
+      "path": "reformer/patterns/iterative_pattern_extraction.py",
+      "class": "IterativePatternExtractor",
+      "methods": [
+        "extract_patterns_from_batch",
+        "run_iterative_extraction"
+      ]
+    }
+  },
+  "profiles": {
+    "rewrite.reformer.repo_native_r_only": {
+      "semantics": "r_only",
+      "pattern_mode": "automatic"
+    },
+    "rewrite.reformer.paper_spec_q_concat_r": {
+      "semantics": "q_concat_r",
+      "pattern_mode": "automatic"
+    },
+    "rewrite.reformer.controlled_r_only": {
+      "semantics": "r_only",
+      "pattern_mode": "manual"
+    },
+    "rewrite.reformer.controlled_q_concat_r": {
+      "semantics": "q_concat_r",
+      "pattern_mode": "manual"
+    }
+  },
+  "execution": {
+    "mode": "user_mounted_official_inference",
+    "vendored_upstream_source": false,
+    "license_acknowledgement_required_at_startup": true,
+    "license_acknowledgement_required_per_request": true,
+    "paper_result_parity_claimed": false,
+    "pattern_induction_runtime_enabled": false,
+    "pattern_induction_symbols_audited": true
+  }
+}