code/annotate_llm.py

"""
Annotate documents using Together AI or Claude Sonnet APIs.
===========================================================
Actions: run, submit-claude, collect-claude, compare.

Usage:
    # Run annotations via Together AI (sync, with optional concurrency)
    python annotate_openrouter.py run \
        --documents ../audit_output/flagged_documents.csv.gz \
        --sample 100000 \
        --model meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8 \
        --output ../audit_output/annotations_llama_100k.csv \
        --label-prefix llama --concurrency 64

    # Submit/collect Claude batch
    python annotate_openrouter.py submit-claude \
        --documents ../audit_output/flagged_documents.csv.gz --sample 300
    python annotate_openrouter.py collect-claude \
        --documents ../audit_output/flagged_documents.csv.gz --sample 300 \
        --batch-id <batch_id> --output ../audit_output/annotations_claude.csv

    # Compare two annotation files
    python annotate_openrouter.py compare \
        --file-a annotations_a.csv --file-b annotations_b.csv

Note: run/collect must use the same --documents, --sample, and seed
to produce matching rows. Uses reservoir sampling for memory efficiency.
"""

import csv
import gzip
import os
import re
import random
import argparse
from collections import Counter

csv.field_size_limit(10_000_000)

# ── Config ────────────────────────────────────────────────────────────────────

LLAMA_MODEL  = "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8"
CLAUDE_MODEL = "claude-sonnet-4-6"

MAX_TEXT_CHARS  = 10000
RANDOM_SEED     = 42

CLASSES = [
    "climate_denial",
    "health_misinfo",
    "pseudoscience",
    "hate_extremism",
    "conspiracy_propaganda",
    "benign",
]

SYSTEM_PROMPT = """You are an expert in misinformation research and content analysis.

Classify whether the following web document contains misinformation, propaganda, or other problematic content. The text may be truncated.

Use exactly one of these labels:
- climate_denial: denies or misrepresents climate change science
- health_misinfo: health or medical misinformation (antivax, COVID denial, quack remedies, unproven supplement claims, alternative medicine promoted as replacement for evidence-based treatment)
- pseudoscience: promotes non-medical unscientific claims (creationism, flat earth, ancient aliens, astrology, etc.)
- hate_extremism: hate speech, Holocaust denial, white nationalism, racist pseudoscience, far-right extremism, or religiously-motivated content that incites hatred or violence against specific groups
- conspiracy_propaganda: conspiracy theories (QAnon, chemtrails, etc.) or state-sponsored propaganda and disinformation
- benign: legitimate content that is not misinformation, propaganda, or pseudoscience

Important: only flag content that actively promotes or endorses misinformation. Use benign for:
- low-quality, politically biased, or opinionated content that is not clearly misinformation
- satirical, humorous, or fictional content
- articles that discuss or critically examine a questionable topic without endorsing it
- content with minor factual errors that is otherwise informational
- speculative or futuristic science and engineering discussions
- historical or archival documents that reflect outdated views or language

Respond in exactly this format:
LABEL: <label>
CONFIDENCE: high | medium | low
REASON: <one sentence explaining your verdict>"""


# ── Data loading ──────────────────────────────────────────────────────────────

def load_documents(documents_path, sample_size, seed=RANDOM_SEED):
    """Load documents from gzipped or plain CSV, optionally sampling.

    Uses reservoir sampling to avoid loading all rows into memory when
    sample_size is set (important for large files like 490K docs).
    """
    is_gz = str(documents_path).endswith(".gz")
    opener = gzip.open if is_gz else open
    mode = "rt" if is_gz else "r"

    def parse_row(row):
        url = row.get("url", "")
        domain = row.get("domain", "")
        if not domain and url:
            try:
                from urllib.parse import urlparse
                domain = urlparse(url).netloc
                if domain.startswith("www."):
                    domain = domain[4:]
            except Exception:
                pass
        return {
            "url":      url,
            "domain":   domain,
            "category": row.get("category", "unknown"),
            "score":    row.get("score", ""),
            "text":     row.get(text_field, "")[:MAX_TEXT_CHARS],
        }

    with opener(documents_path, mode) as f:
        reader = csv.DictReader(f)
        fields = reader.fieldnames
        text_field = "full_text" if "full_text" in fields else "text_preview"

        if not sample_size:
            rows = [parse_row(row) for row in reader]
            print(f"  Loaded {len(rows)} documents")
            return rows

        # Reservoir sampling: keeps only sample_size rows in memory
        rng = random.Random(seed)
        reservoir = []
        total = 0
        for i, row in enumerate(reader):
            total = i + 1
            if i < sample_size:
                reservoir.append(parse_row(row))
            else:
                j = rng.randint(0, i)
                if j < sample_size:
                    reservoir[j] = parse_row(row)

        print(f"  Sampled {len(reservoir)} from {total:,} total (reservoir sampling)")
        return reservoir


# ── Prompt ────────────────────────────────────────────────────────────────────

def make_prompt(row):
    return f"""Document text:
{row['text']}

Classify this document."""


def strip_thinking(text):
    """Remove <think>...</think> blocks that some models may produce."""
    return re.sub(r"<think>.*?</think>", "", text, flags=re.DOTALL).strip()


def parse_response(text):
    """Extract label, confidence, reason from model response."""
    text = strip_thinking(text)
    label = confidence = reason = ""
    for line in text.strip().splitlines():
        line = line.strip()
        if line.upper().startswith("LABEL:"):
            label = line.split(":", 1)[1].strip().lower()
        elif line.upper().startswith("CONFIDENCE:"):
            confidence = line.split(":", 1)[1].strip().lower()
        elif line.upper().startswith("REASON:"):
            reason = line.split(":", 1)[1].strip()
    if label not in CLASSES:
        label = "unknown"
    return label, confidence, reason


# ── Claude batch ──────────────────────────────────────────────────────────────

def submit_claude_batch(rows):
    import anthropic
    from anthropic.types.message_create_params import MessageCreateParamsNonStreaming
    from anthropic.types.messages.batch_create_params import Request

    client = anthropic.Anthropic()
    requests = [
        Request(
            custom_id=f"item-{i}",
            params=MessageCreateParamsNonStreaming(
                model=CLAUDE_MODEL,
                max_tokens=128,
                temperature=0.1,
                system=SYSTEM_PROMPT,
                messages=[{"role": "user", "content": make_prompt(row)}],
            )
        )
        for i, row in enumerate(rows)
    ]
    batch = client.messages.batches.create(requests=requests)
    print(f"  Batch submitted: {batch.id}  ({len(requests)} requests)")
    return batch.id


def collect_claude_batch(batch_id, n_rows):
    import anthropic
    client = anthropic.Anthropic()

    batch = client.messages.batches.retrieve(batch_id)
    print(f"  Batch status: {batch.processing_status}")
    if batch.processing_status != "ended":
        print(f"  ERROR: batch not complete")
        return None

    results = [""] * n_rows
    for result in client.messages.batches.results(batch_id):
        idx = int(result.custom_id.split("-")[1])
        if result.result.type == "succeeded":
            msg = result.result.message
            results[idx] = next((b.text for b in msg.content if b.type == "text"), "")
        else:
            results[idx] = f"ERROR: {result.result.type}"

    succeeded = sum(1 for r in results if r and not r.startswith("ERROR"))
    print(f"  Collected {succeeded}/{n_rows} results")
    return results


# ── Synchronous / Concurrent Together AI ─────────────────────────────────────

def run_together_sync(rows, model, output_path, label_prefix="llm", concurrency=1):
    """Run Together AI requests, optionally with concurrency. Saves incrementally."""
    checkpoint_path = output_path + ".checkpoint"
    p = f"{label_prefix}_" if label_prefix else ""
    fieldnames = ["url", "domain", "category", "score",
                  f"{p}label", f"{p}confidence", f"{p}reason"]

    # Resume from checkpoint if exists
    completed = {}
    if os.path.exists(checkpoint_path):
        with open(checkpoint_path) as f:
            reader = csv.DictReader(f)
            for row in reader:
                completed[row["url"]] = row
        print(f"  Resuming from checkpoint: {len(completed):,} already done")

    # Filter out already-completed rows
    remaining_indices = [i for i, row in enumerate(rows) if row["url"] not in completed]
    print(f"  {len(remaining_indices):,} remaining to annotate")

    if remaining_indices:
        if concurrency > 1:
            responses = _run_concurrent_incremental(
                rows, remaining_indices, model, concurrency,
                checkpoint_path, fieldnames, p)
        else:
            responses = [""] * len(rows)
            # Run remaining
            from together import Together
            client = Together()
            for idx in remaining_indices:
                row = rows[idx]
                try:
                    resp = client.chat.completions.create(
                        model=model,
                        messages=[
                            {"role": "system", "content": SYSTEM_PROMPT},
                            {"role": "user",   "content": make_prompt(row)},
                        ],
                        max_tokens=256, temperature=0.1,
                    )
                    responses[idx] = resp.choices[0].message.content
                except Exception as e:
                    responses[idx] = f"ERROR: {e}"
                # Append to checkpoint
                label, conf, reason = parse_response(responses[idx])
                with open(checkpoint_path, "a", newline="") as f:
                    writer = csv.DictWriter(f, fieldnames=fieldnames)
                    if f.tell() == 0:
                        writer.writeheader()
                    writer.writerow({
                        "url": row["url"], "domain": row["domain"],
                        "category": row["category"], "score": row["score"],
                        f"{p}label": label, f"{p}confidence": conf, f"{p}reason": reason,
                    })

        # Fill in already-completed responses from checkpoint
        for i, row in enumerate(rows):
            if row["url"] in completed:
                c = completed[row["url"]]
                responses[i] = f"LABEL: {c[f'{p}label']}\nCONFIDENCE: {c[f'{p}confidence']}\nREASON: {c[f'{p}reason']}"

        save_annotations(rows, responses, output_path, label_prefix=label_prefix)
        return

    # Build full responses from checkpoint + new results for final save
    all_responses = [""] * len(rows)
    for i, row in enumerate(rows):
        if row["url"] in completed:
            c = completed[row["url"]]
            all_responses[i] = f"LABEL: {c[f'{p}label']}\nCONFIDENCE: {c[f'{p}confidence']}\nREASON: {c[f'{p}reason']}"

    save_annotations(rows, all_responses, output_path, label_prefix=label_prefix)


def _run_concurrent_incremental(rows, indices, model, concurrency,
                                 checkpoint_path=None, fieldnames=None, prefix=""):
    """Run requests with asyncio concurrency, saving to checkpoint incrementally."""
    import asyncio
    from openai import AsyncOpenAI

    client = AsyncOpenAI(
        api_key=os.environ["TOGETHER_API_KEY"],
        base_url="https://api.together.xyz/v1",
    )

    responses = [""] * len(rows)
    done_count = 0
    lock = asyncio.Lock() if checkpoint_path else None

    # Open checkpoint file for appending
    cp_file = None
    cp_writer = None
    if checkpoint_path:
        needs_header = not os.path.exists(checkpoint_path) or os.path.getsize(checkpoint_path) == 0
        cp_file = open(checkpoint_path, "a", newline="")
        cp_writer = csv.DictWriter(cp_file, fieldnames=fieldnames)
        if needs_header:
            cp_writer.writeheader()
            cp_file.flush()

    async def process(i, row, sem):
        nonlocal done_count
        async with sem:
            try:
                resp = await client.chat.completions.create(
                    model=model,
                    messages=[
                        {"role": "system", "content": SYSTEM_PROMPT},
                        {"role": "user",   "content": make_prompt(row)},
                    ],
                    max_tokens=256,
                    temperature=0.1,
                )
                responses[i] = resp.choices[0].message.content
            except Exception as e:
                responses[i] = f"ERROR: {e}"

            # Write to checkpoint
            if cp_writer and lock:
                label, conf, reason = parse_response(responses[i])
                async with lock:
                    cp_writer.writerow({
                        "url": row["url"], "domain": row["domain"],
                        "category": row["category"], "score": row["score"],
                        f"{prefix}label": label, f"{prefix}confidence": conf,
                        f"{prefix}reason": reason,
                    })
                    done_count += 1
                    if done_count % 1000 == 0:
                        cp_file.flush()
                        print(f"  {done_count}/{len(indices)} done (checkpoint saved)", flush=True)
                    elif done_count % 100 == 0 or done_count == len(indices):
                        print(f"  {done_count}/{len(indices)} done", flush=True)
            else:
                done_count += 1
                if done_count % 100 == 0 or done_count == len(indices):
                    print(f"  {done_count}/{len(indices)} done", flush=True)

    async def run_all():
        sem = asyncio.Semaphore(concurrency)
        tasks = [process(idx, rows[idx], sem) for idx in indices]
        await asyncio.gather(*tasks)

    asyncio.run(run_all())

    if cp_file:
        cp_file.flush()
        cp_file.close()

    return responses


# ── Save results ──────────────────────────────────────────────────────────────

def save_annotations(rows, responses, output_path, label_prefix=""):
    """Save annotations to CSV."""
    p = f"{label_prefix}_" if label_prefix else ""
    fieldnames = ["url", "domain", "category", "score",
                  f"{p}label", f"{p}confidence", f"{p}reason"]

    with open(output_path, "w", newline="") as f:
        writer = csv.DictWriter(f, fieldnames=fieldnames)
        writer.writeheader()
        for i, row in enumerate(rows):
            label, conf, reason = parse_response(responses[i])
            writer.writerow({
                "url":             row["url"],
                "domain":          row["domain"],
                "category":        row["category"],
                "score":           row["score"],
                f"{p}label":      label,
                f"{p}confidence": conf,
                f"{p}reason":     reason,
            })

    print(f"\nSaved {len(rows)} annotations to {output_path}")

    # Summary
    labels = [parse_response(r)[0] for r in responses]
    print(f"\nLabel distribution:")
    for label, n in Counter(labels).most_common():
        print(f"  {label:<25}: {n:4d} ({n/len(rows)*100:.1f}%)")


# ── Compare ───────────────────────────────────────────────────────────────────

def compare_annotations(file_a, file_b, name_a="A", name_b="B"):
    """Compare two annotation files and print agreement stats."""
    rows_a = list(csv.DictReader(open(file_a)))
    rows_b = list(csv.DictReader(open(file_b)))

    # Detect label column names
    def find_label_col(fieldnames):
        for f in fieldnames:
            if f.endswith("label"):
                return f
        return None

    label_col_a = find_label_col(rows_a[0].keys()) if rows_a else None
    label_col_b = find_label_col(rows_b[0].keys()) if rows_b else None

    if not label_col_a or not label_col_b:
        print("ERROR: could not find label columns")
        return

    # Build URL→label maps
    map_a = {r["url"]: r[label_col_a] for r in rows_a}
    map_b = {r["url"]: r[label_col_b] for r in rows_b}

    common_urls = set(map_a.keys()) & set(map_b.keys())
    print(f"\n{name_a}: {len(rows_a)} docs ({label_col_a})")
    print(f"{name_b}: {len(rows_b)} docs ({label_col_b})")
    print(f"Common URLs: {len(common_urls)}")

    if not common_urls:
        print("ERROR: no common URLs to compare")
        return

    agreements = sum(1 for u in common_urls if map_a[u] == map_b[u])
    print(f"\nAgreement: {agreements}/{len(common_urls)} ({agreements/len(common_urls)*100:.1f}%)")

    # Disagreements
    disagree = [(map_a[u], map_b[u]) for u in common_urls if map_a[u] != map_b[u]]
    if disagree:
        print(f"\nDisagreements ({name_a} → {name_b}):")
        for pair, count in Counter(disagree).most_common(15):
            print(f"  {pair[0]:<25} → {pair[1]:<25} ({count})")

    # Per-class agreement
    print(f"\nPer-class agreement:")
    for cls in CLASSES + ["unknown"]:
        in_a = [u for u in common_urls if map_a[u] == cls]
        if in_a:
            agree = sum(1 for u in in_a if map_b[u] == cls)
            print(f"  {cls:<25}: {agree}/{len(in_a)} ({agree/len(in_a)*100:.0f}%)")


# ── Main ──────────────────────────────────────────────────────────────────────

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    subparsers = parser.add_subparsers(dest="action", required=True)

    # ── run (Together AI, sync or concurrent)
    p_run = subparsers.add_parser("run", help="Run Together AI annotations (sync or concurrent)")
    p_run.add_argument("--documents", required=True)
    p_run.add_argument("--sample", type=int, default=300)
    p_run.add_argument("--model", default=LLAMA_MODEL, help=f"Model name (default: {LLAMA_MODEL})")
    p_run.add_argument("--output", required=True)
    p_run.add_argument("--label-prefix", default="llm", help="Column prefix (default: llm)")
    p_run.add_argument("--concurrency", type=int, default=1, help="Number of concurrent requests (default: 1)")

    # ── submit-claude
    p_submit_c = subparsers.add_parser("submit-claude", help="Submit Claude batch")
    p_submit_c.add_argument("--documents", required=True)
    p_submit_c.add_argument("--sample", type=int, default=300)

    # ── collect-claude
    p_collect_c = subparsers.add_parser("collect-claude", help="Collect Claude results")
    p_collect_c.add_argument("--documents", required=True)
    p_collect_c.add_argument("--sample", type=int, default=300)
    p_collect_c.add_argument("--batch-id", required=True)
    p_collect_c.add_argument("--output", required=True)

    # ── compare
    p_compare = subparsers.add_parser("compare", help="Compare two annotation files")
    p_compare.add_argument("--file-a", required=True)
    p_compare.add_argument("--file-b", required=True)
    p_compare.add_argument("--name-a", default="A")
    p_compare.add_argument("--name-b", default="B")

    args = parser.parse_args()

    if args.action == "run":
        print(f"Loading documents...")
        rows = load_documents(args.documents, args.sample)
        print(f"\nRunning {len(rows)} requests (model: {args.model}, concurrency: {args.concurrency})...")
        run_together_sync(rows, args.model, args.output, label_prefix=args.label_prefix, concurrency=args.concurrency)

    elif args.action == "submit-claude":
        print("Loading documents...")
        rows = load_documents(args.documents, args.sample)
        print(f"\nSubmitting Claude batch...")
        submit_claude_batch(rows)

    elif args.action == "collect-claude":
        print("Loading documents...")
        rows = load_documents(args.documents, args.sample)
        print(f"\nCollecting Claude batch {args.batch_id}...")
        responses = collect_claude_batch(args.batch_id, len(rows))
        if responses:
            save_annotations(rows, responses, args.output, label_prefix="claude")

    elif args.action == "compare":
        compare_annotations(args.file_a, args.file_b, args.name_a, args.name_b)