Added all the files required to run task generation agentic pipeline#59
Added all the files required to run task generation agentic pipeline#59saidul-islam98 wants to merge 1 commit intomainfrom
Conversation
kohankhaki
left a comment
kohankhaki
left a comment
There was a problem hiding this comment.
Thanks for the new agentic work. I have one question, is this going to replace the stage 3 in agentic workflow? Or introduces an alternative for the stage 3?
From the current repo structure, we have two paths:
- The schema-standard base pipeline (
src/run_base_pipeline.py+src/base_stages/*, documented insrc/schemas/GENERATION_PIPELINE_SCHEMAS.md). - A legacy agentic debate path (
src/agentic_*entrypoints) that uses older custom JSON structures.
For this PR, since the goal is to make the new agentic Stage 3 replaceable with the current base Stage 3, the output contract needs to match the standardized schema exactly (Stage-3 layout, metadata linkage, hierarchy fields, ID conventions), so Stage 4/5 can consume it without special handling.
Stage 3 should consume capabilities/<capabilities_tag>/<area_id>/capabilities.json in the standardized format.
To test this Stage-3 implementation, please use schema-standard Stage-2 inputs:
- either generate areas/capabilities via the standard pipeline (stages 0–2), or
- create custom areas/capabilities using the standard dataclasses (
Domain,Area,Capability) andsave_capabilities.
| ) | ||
|
|
||
| # chapter path for output | ||
| chapter_out_path = ( |
There was a problem hiding this comment.
Output path is non-standard (tasks/<tag>/<book>/<chapter>/tasks.json). Standard schema requires tasks/<task_tag>/<area_id>/<capability_id>/tasks.json. Please align path structure to schema contract.
| def make_verifier_agent() -> VerifierAgent: | ||
| return VerifierAgent(name="Verifier", model_client=verifier_client) | ||
|
|
||
| for chapter_idx, chapter_path in enumerate(chapter_files): |
There was a problem hiding this comment.
Stage-3 standard flow should iterate over Stage-2 capabilities (area/capability hierarchy), not chapter files directly. This currently bypasses standardized Stage-2 -> Stage-3 contract and breaks immediate Stage-4/5 interoperability.
| dedup_cfg.get("embedding_model", "text-embedding-3-small") | ||
| ) | ||
| keep_policy = str(dedup_cfg.get("keep_policy", "first")) | ||
| cache_embeddings = bool(dedup_cfg.get("cache_embeddings", True)) |
There was a problem hiding this comment.
input_stage_tag is set to None. For Stage-3 this must reference the Stage-2 capabilities tag for provenance/resume compatibility. Please pass the actual input tag.
| ) | ||
| continue | ||
|
|
||
| all_tasks: List[Task] = [] |
There was a problem hiding this comment.
Tasks are being built with placeholder capability/area/domain fields (__placeholder__, *_placeholder).
Schema-compliant Task objects must include real hierarchy values from actual Capability inputs (capability/area/domain identifiers and names), not placeholders.
| meta["chapter_id"] = meta.get("chapter_id") or chapter_id | ||
| t.generation_metadata = meta | ||
|
|
||
| t.task_id = f"{prefix}__task_{i:03d}" |
There was a problem hiding this comment.
Dedup rewrites task_id to <chapter_id>__task_###, which breaks standardized task ID format and scope expectations.
Please keep/assign schema-standard task_### IDs in capability scope after deduplication as well.
2 participants
PR Type
Feature
Short Description
Added the required files to run the task generation agentic pipeline. The instructions to run the pipeline can be found under:
Tests Added
None