Social Media Sentiment Analysis: A Data Pipeline That Starts With Clean Records

Quickstart

Every request needs one header: Authorization: Bearer $FETCHPI_API_KEY. Create a free key in the console and you can replay every snippet on this page as-is.

curl -s -H "Authorization: Bearer $FETCHPI_API_KEY"   "https://fetchpi.com/v1/x/search/posts?q=your-topic&freshness=default"

Sentiment quality is data quality

Sentiment models amplify input noise. Ad entries, deleted authors, unparsable timestamps and layout artifacts all become training/prediction poison. FetchPI filters at the extraction layer — promoted entries are dropped, timestamps are normalized to RFC 3339, authors are typed records — so the frame you hand to a model is already clean.

The pipeline

• Collect: keyword search per platform (freshness=default, max_age you can afford).
• Normalize: records already share one schema — dedupe on id, bucket on created_at.
• Score: your model (or an LLM pass) labels each text.
• Serve: aggregate by platform/subreddit/author; every number traces to meta.request_id lineage.

records = []
cursor = None
while True:
    params = {"q": "your-brand", "freshness": "default", "max_age": "900"}
    if cursor:
        params["cursor"] = cursor
    r = requests.get(
        "https://fetchpi.com/v1/reddit/search/posts",
        headers={"Authorization": f"Bearer {os.environ['FETCHPI_API_KEY']}"},
        params=params, timeout=30,
    )
    page = r.json()
    records += page["data"]
    cursor = page.get("next_cursor")
    if not cursor:
        break
# records: typed, dedupeable, ready for scoring

Cost shape

Collection dominates the bill. Cache-mode reads bill at the catalog cache tier (most capabilities half price) and failed upstream calls are never billed — the pipeline's unit economics survive the retry storms that kill scraper-based sentiment projects.