# How do I cut the tokens when I feed social media API results to an LLM?

> A page of Douyin or Xiaohongshu posts in full JSON fills thousands of tokens. Send the model only the fields the task needs, at the same price per call.

Canonical page: https://okdata.online/answers/reduce-llm-tokens-social-data/

**Short answer: **Ask the API for less instead of trimming afterwards: okdata takes a list of the fields to keep and a compact view that drops empty values, and neither changes the price ($0.02 per Xiaohongshu search page either way). For a comment sentiment task, three fields per comment is usually all the model needs.

Three Douyin searches into a research session, the context is half full and the agent starts forgetting the brief. Claude Code may even move a large tool result to a file because it is too big to show. The fix is to ask for less data per item, not fewer items: a page of 20 videos where each video carries five fields is a small fraction of the same page with author avatars, image lists, hashtags and every metric.

## Two parameters that shrink every response

Every okdata endpoint, from [Douyin video search](/apis/douyin/videos/search/) to [Xiaohongshu note search](/apis/xiaohongshu/notes/search/) and [Douyin video comments](/apis/douyin/videos/comments/list/), takes the same two options:

- `fields` keeps only the paths you list, with dots for nested values: `fields=id,title,metrics.likes`. A path that matches nothing is reported in `warnings`, so typos do not fail silently.
- `view` sets how much comes back by default. `standard` is the default, `compact` also drops every null field, and `full` adds platform extras under `ext` that you rarely want in a prompt.

Neither changes the price or the number of items. You pay per call, so the slim page costs the same as the fat one.

## Which fields for which job

| Task | fields | Price per page |
|---|---|---|
| Comment sentiment | `text,likes,published_at` | $0.002 |
| Hot list scan | `rank,name,heat` | $0.002 |
| Creator shortlist | `id,name,followers,posts,likes,bio` | $0.002 |
| Content research | `id,url,title,text,metrics.likes,metrics.comments,author.name` | $0.02 |

## See the difference yourself

This runs the same search twice, once in the default shape and once trimmed, and prints the size of each. It is two calls, $0.04 in total.

```python
import json, os, requests

URL = "https://api.okdata.online/v1/douyin/videos/search"
H = {"Authorization": f"Bearer {os.environ['OKDATA_API_KEY']}"}

for fields in (None, "id,title,metrics.likes,author.name"):
    params = {"query": "露营装备", **({"fields": fields} if fields else {})}
    body = requests.get(URL, headers=H, params=params, timeout=60).json()
    if body["code"] != "ok":
        raise RuntimeError(body["message"])
    size = len(json.dumps(body["data"], ensure_ascii=False))
    print(fields or "standard", size, "characters", body["warnings"])
```

Over MCP, the agent passes the same thing inside `params`, for example `{"query": "露营装备", "fields": "id,title,metrics.likes"}`. Tell it to do so in your prompt; models rarely ask for fewer fields on their own.

## Habits that save more than any parameter

- **Keep the raw data, show the model the slim copy.** Save full responses to a file or database, and pass the model only what the current step needs. You can always go back to the raw data without paying again.
- **Summarize each page before fetching the next.** Ten pages of comments become ten short notes, then one summary, instead of one giant prompt.
- **Transcripts:** `fields=text` returns the plain transcript without per-segment timestamps, often half the size. Keep the segments only when you need to cite the second.
- **Cap the items, not just the fields.** If the model only needs the top 20 comments by likes, sort in code and drop the rest before the prompt.

## Other ways to do it

Post-processing with `jq` or a few lines of Python works with any API and costs nothing; it just means the full response travels to your code first. Letting the MCP client truncate long results is the worst option: it cuts from the end, so the model silently loses the last items rather than the useless fields. The full parameter list is on the [requests page](/docs/requests/).

## Endpoints used

- [Xiaohongshu note search](https://okdata.online/apis/xiaohongshu/notes/search/) `xiaohongshu.notes.search`: $0.02 per call · about 20 notes per page
- [Douyin video search](https://okdata.online/apis/douyin/videos/search/) `douyin.videos.search`: $0.02 per call · about 20 videos per page
- [Douyin video comments](https://okdata.online/apis/douyin/videos/comments/list/) `douyin.videos.comments.list`: $0.002 per call · about 20 comments per page

Updated 2026-11-12