BurnyCoder commited on
Commit
f1eda77
·
verified ·
1 Parent(s): 0c0d808

Update research card

Browse files
Files changed (1) hide show
  1. README.md +16 -40
README.md CHANGED
@@ -1,40 +1,16 @@
1
- # Luna folder research swarm
2
-
3
- A private experiment in answering questions from files. Independent GPT-6 Luna investigators use Codex to search a read-only folder. A coordinator checks their source excerpts and writes a cited answer. The experiment varies the total number of agents and measures answer quality, evidence support, time, and API cost on news and Python repositories.
4
-
5
- The starting observation comes from a small, model-specific study. Its single agent outscored its swarm on repository questions. Its filename-recall measure leaves open whether an agent saw a passage that substantiates the answer. The present comparison holds the generation model fixed, includes ordinary Codex and one-shot retrieval, and measures passage delivery. [Arctic results](https://github.com/BurnyCoder/arcticswarm-folder-rag-research/blob/bd5b3ed98cde10e9c1f90795f2182d84ecd51ddc/results/day1/tables.md) · [Original collector](https://github.com/BurnyCoder/arcticswarm-folder-rag-research/blob/bd5b3ed98cde10e9c1f90795f2182d84ecd51ddc/src/folderrag/collect.py#L78)
6
-
7
- ## Research record
8
-
9
- - [Literature review](reports/literature-review.md) and [state of the art](reports/state-of-the-art-review.md)
10
- - [Preregistration](reports/preregistration.md) and [implementation plan](reports/plan.md)
11
-
12
- ## Reproduction
13
-
14
- The runtime is Linux with bubblewrap, ripgrep, Codex CLI, Python 3.12 and uv. Windows users can run it with WSL2. A GPU accelerates the local Qwen embedding encoder. Keep the OpenAI API key in an ignored project-root `.env`. The host process handles credentials. Each research arm receives only its question and permitted corpus. The budget covers all paid inference and grading for this study. [Codex CLI](https://learn.chatgpt.com/docs/non-interactive-mode) · [Luna](https://developers.openai.com/api/docs/models/gpt-6-luna) · [Qwen model card](https://huggingface.co/Qwen/Qwen3-Embedding-0.6B)
15
-
16
- Run `uv run folderrag ask --folder PATH --agents 4 "QUESTION"`. Choose 1, 2, 4 or 8 total agents. At size four, three independent investigators gather evidence and one coordinator verifies quotations and produces the answer. Add `--json` for the answer, citations, offsets, worker completions, time and cost.
17
-
18
- ![Luna investigator and coordinator architecture](reports/figures/architecture.png)
19
-
20
- [Installation, live dashboard and reproduction commands](docs/usage.md) · [Frozen question manifest](protocol/splits.json) · [Private results and examples](https://huggingface.co/spaces/BurnyCoder/luna-folder-swarm-explorer)
21
-
22
- ## Research milestones
23
-
24
- - [x] Register the protocol and ordered, stratified evaluation design.
25
- - [x] Implement the swarm, four sizes, baselines, cost ledger, and source validation.
26
- - [ ] Complete the development sweep and freeze the candidate size and affordable test set.
27
- - [ ] Complete the paired confirmation, controls, intervention tests and evidence audit.
28
-
29
- <!-- generated -->
30
- ## Measured results
31
-
32
- The protocol and implementation are available. Paired experiments are pending.
33
-
34
- | Phase | Domain | Method | Graded / attempts | Quality [95% CI] | USD/question¹ | Median / p95 s | Errors |
35
- |---|---|---|---:|---:|---:|---:|---:|
36
- | Pending | | | | | | | |
37
-
38
- ¹ Conservative inference charge using the cache-write upper rate for unclassified input tokens. Grading is included in the phase ledger.
39
-
40
- [Paper](reports/paper.pdf) · [All experiment reports](reports/summary.md)
 
1
+ ---
2
+ language:
3
+ - en
4
+ task_categories:
5
+ - question-answering
6
+ tags:
7
+ - retrieval-augmented-generation
8
+ - multi-agent
9
+ - evaluation
10
+ ---
11
+
12
+ # Luna folder swarm research
13
+
14
+ Private research records for a preregistered comparison of single and multi-agent file search. Each condition uses GPT-6 Luna. Codex investigators search an isolated folder and send evidence to a coordinator. The experiment varies total agent count and includes stock Codex, Arctic B2, one-shot retrieval, controls and ablations.
15
+
16
+ See the [preregistration](reports/preregistration.md), [measured tables](results/tables.md), [paper](reports/paper.pdf) and [private results explorer](https://huggingface.co/spaces/BurnyCoder/luna-folder-swarm-explorer). The [private GitHub repository](https://github.com/BurnyCoder/agent-swarm-for-retrieval-augmented-generation) contains the exact implementation and pinned Arctic submodule. The source archive contains Git-tracked files. Benchmark sources retain their individual licenses. The case records distinguish development, confirmation and diagnostics.