RescueLens: LLM Triage for Food Rescue

• RescueLens is an LLM-powered system designed to triage volunteer feedback in food rescue operations by categorizing free-text comments into actionable issues.
• It employs prompt-based in-context learning with models like GPT-4o mini to classify feedback into seven issue types and rank donors and recipients based on issue frequencies.
• Evaluated on data from 412 Food Rescue, it achieves up to 96% recall and significantly reduces manual triage time from hours per week to minutes per day.

Searching arXiv for the named system and closely related work so the article can be grounded in current papers and the provided primary source. RescueLens is an LLM-powered tool for triage and action on volunteer feedback in food rescue operations. Developed with 412 Food Rescue, a large food rescue organization based in Pittsburgh, Pennsylvania, it is designed to automatically categorize volunteer feedback, suggest donors and recipients for follow-up, and update volunteer directions when comments indicate outdated or incorrect pickup or drop-off instructions. The system addresses a setting in which free-text feedback is abundant but manual monitoring is cumbersome and labor-intensive, and it has been evaluated on annotated organizational data and deployed in production at 412 Food Rescue (Raman et al., 19 Nov 2025).

1. Organizational setting and problem formulation

Food rescue organizations redistribute food from donors who have excess to recipients who need it, while relying heavily on volunteers to execute rescue trips. In the case studied by RescueLens, 412 Food Rescue receives approximately 1,800 free-text feedbacks per year, yet historically had no standardized way to log, categorize, and prioritize them. Organizers therefore spent hours each week reading comments one-by-one, with little visibility into which issues were most frequent or severe, and the lack of formal documentation meant that issues could “fall through the cracks” or remain difficult to distinguish as novel versus recurring (Raman et al., 19 Nov 2025).

RescueLens formalizes this operational problem into three high-level goals. First, it performs automatic categorization of volunteer comments into seven actionable issue types. Second, it identifies and ranks donors and recipients whose rescue trips yield high issue rates, enabling targeted organizer outreach. Third, it automates rewriting of volunteer directions when comments explicitly indicate direction errors or omissions. In this formulation, volunteer feedback is treated not merely as qualitative commentary, but as an operational signal for triage, prioritization, and intervention.

A central design choice is that the system is optimized for issue recovery rather than conservative filtering. The paper explicitly characterizes RescueLens as “high-recall, moderate-precision” by design: organizers can filter false positives, but cannot afford to miss real issues. This design principle shapes both the evaluation framing and the deployment logic.

2. Data sources and annotation schema

The underlying dataset comes from 412 Food Rescue’s production database of post-trip feedback spanning 2018–2025. The organizational scale reported in the paper is substantial: more than 200,000 recorded rescue trips and 14,439 unique free-text feedback entries (Raman et al., 19 Nov 2025).

The annotation schema was developed through open-coding on 250 randomly sampled 2024 comments and refined until thematic exhaustion. Inter-annotator reliability was measured on 125 held-out 2024 comments, yielding Cohen’s $\kappa = 0.73$, reported as significant agreement. The resulting taxonomy comprises seven issue categories:

Category
Inadequate Food
Earlier Pickup
Donor Problem
Recipient Problem
Update Contact
System Problem
Direction Problem

The annotation guidelines include precise “label vs. no-label” rules. One example given in the summary is that “Earlier Pickup” is not equivalent to “Inadequate Food.” This matters because the categories are explicitly operational rather than merely topical: the labels are intended to support organizer action, donor or recipient follow-up, or direction maintenance.

The taxonomy also defines the scope of the downstream ranking and rewriting modules. A comment may map to one or more categories, and that multi-label structure is handled through prompt output in JSON. The dependence on an organization-specific coding scheme is important: the paper states that prompts and taxonomy were tuned to 412 Food Rescue, and that transfer to other nonprofits will require re-coding categories and guidelines.

3. Model architecture and prompt design

RescueLens uses GPT-4o mini with in-context learning and no fine-tuning as its default model. Comparative LLMs include GPT-4o, Llama 3.1, and DeepSeek R1, alongside non-LLM baselines TF-IDF+LR and DistilBERT. Because the system uses few-shot in-context learning, there is no explicit fine-tuning or cross-entropy loss; classification is prompt-based rather than parameter-updated (Raman et al., 19 Nov 2025).

Each category task prompt contains three elements: a brief scenario description, detailed bullet-point labeling guidelines, and 3–8 few-shot examples with explanations. The prompt design therefore encodes task semantics procedurally rather than through supervised weight adaptation. Operationally, free text is passed to the model, and the model returns one or more of the seven categories in JSON.

This architecture is notable for the absence of a custom-trained classifier despite the availability of annotated data. A plausible implication is that the designers prioritized rapid deployment, prompt auditability, and low engineering overhead over the development of a task-specific fine-tuned model. The paper itself does not frame this as a universal superiority claim; rather, it presents GPT-4o mini as the default within a comparative evaluation.

The same prompt-centric design extends to the direction rewrite module. When feedback is labeled “Direction Problem,” the model is prompted with the original directions plus the volunteer comment and instructed to only add or correct explicitly stated information, not delete existing details, and preserve all unchanged instructions. The output is structured JSON containing booleans for donor_direction_change and recipient_direction_change, along with rewritten strings.

4. Triage logic, ranking functions, and direction updates

The automatic categorization component is evaluated using standard classification metrics:

$$\text{Precision} = \frac{TP}{TP + FP}, \qquad \text{Recall} = \frac{TP}{TP + FN}, \qquad F1 = \frac{2 \cdot \text{Precision} \cdot \text{Recall}}{\text{Precision} + \text{Recall}}.$$

Within the system, categorization is not an endpoint. Its outputs feed two action modules: entity ranking and direction maintenance. For each donor or recipient entity $e$, RescueLens defines

$$\text{Trips}(e) = \text{total number of trips for } e$$

$$\text{Issues}(e) = \# \text{ trips where } (\text{rating} < 4) \text{ OR } (\text{any issue detected})$$

$$\text{Score}(e) = \frac{\text{Issues}(e)}{\text{Trips}(e)}.$$

Donors and recipients are then sorted in descending $\text{Score}(e)$, and organizers focus on the top-$k$, with the paper using top 0.5% as an illustrative operating point. This converts free-text monitoring into a ranking problem over operational entities rather than a flat stream of comments (Raman et al., 19 Nov 2025).

The direction rewrite module is narrower in scope. It is triggered only when feedback is labeled “Direction Problem,” and the prompt includes explicit decision heuristics. The system updates directions only if the volunteer explicitly reports a wrong phone number, wrong address, or missing entrance detail. It does not update for general complaints, food-quantity issues, or timing delays. This constrained rewrite policy limits the module to directly evidenced corrections rather than open-ended summarization.

Taken together, these components define RescueLens as a triage-and-action pipeline rather than a pure text classifier. Categorization identifies problem types, ranking prioritizes organizational outreach, and rewriting produces candidate operational updates.

5. Evaluation and quantitative findings

The paper reports multiple performance summaries for issue detection. In the abstract, RescueLens is said to recover 96% of volunteer issues at 71% precision. In the detailed summary, for “Any Issue” detection using RescueLens with GPT-4o mini, the reported metrics are Accuracy $= 93.1\%$, Precision $= 83.3\%$, Recall $$\text{Precision} = \frac{TP}{TP + FP}, \qquad \text{Recall} = \frac{TP}{TP + FN}, \qquad F1 = \frac{2 \cdot \text{Precision} \cdot \text{Recall}}{\text{Precision} + \text{Recall}}.$$0, and $$\text{Precision} = \frac{TP}{TP + FP}, \qquad \text{Recall} = \frac{TP}{TP + FN}, \qquad F1 = \frac{2 \cdot \text{Precision} \cdot \text{Recall}}{\text{Precision} + \text{Recall}}.$$1, with cost approximately \$15 per year (Raman et al., 19 Nov 2025).

Per-category results are also reported. Every category achieves at least 75% recall and at least 93% accuracy. The summary highlights “Direction Problem” as having 100% recall, 75.4% precision, and $$\text{Precision} = \frac{TP}{TP + FP}, \qquad \text{Recall} = \frac{TP}{TP + FN}, \qquad F1 = \frac{2 \cdot \text{Precision} \cdot \text{Recall}}{\text{Precision} + \text{Recall}}.$$2. These numbers are consistent with the system’s stated emphasis on recall: false positives are tolerated to reduce missed issues.

The entity-ranking results are operationally striking. Figure 1 is summarized as showing that only 0.5% of donors account for more than 30% of all reported issues. An additional example states that the five worst donors represent 2.5% of trips but cause 25% of issues in each of four major categories. This indicates that the ranking objective surfaces a highly concentrated distribution of operational problems, enabling targeted intervention rather than uniform review.

The evaluation also includes organizer-facing utility. Through ranking donors and recipients according to their rates of volunteer issues, RescueLens enables prioritization of outreach to a small subset of problematic entities. This suggests that much of the organization’s remediation workload can be concentrated on a narrow frontier of high-issue donors and recipients.

6. Deployment, impact, and scope of generalization

RescueLens has been deployed at 412 Food Rescue since May 15, 2025. In production, a daily Ruby script ingests new feedback, runs RescueLens, and writes outputs to a “Rescue Feedback” database table. Monthly reports are sent to Slack and include top donors and recipients by issue score and newly rewritten directions for review. Organizers also use a web UI that allows filtering by date and category and supports leaving notes (Raman et al., 19 Nov 2025).

Semi-structured interviews with organizers provide evidence about day-to-day use. One organizer statement quoted in the summary is: “We look at RescueLens every day to track trends in store leads or ordering changes.” Organizers also report that approximately 50% of flagged comments are actionable and that, before RescueLens, many actionable issues went unnoticed. Reported time savings are from hours per week to minutes per day for triaging feedback, with the additional benefit of enabling quarterly and seasonal trend analysis.

The paper also states several limitations. Volunteer comments can be highly ambiguous, and some corrective suggestions may still slip through. The system does not include built-in tracking of whether an organizer actually intervened after suggestions. Prompts and taxonomy are tuned to 412 Food Rescue, so transfer to other nonprofits requires re-coding categories and guidelines. Proposed extensions include adding a follow-up tracker to log outreach actions and measure downstream impact on issue scores, building a Q&A chatbot for volunteers using the historical feedback corpus, and adapting the taxonomy for other volunteer-driven sectors such as disaster relief and tutoring programs.

Open questions remain around how best to measure the long-term effect of automated triage on volunteer retention and how to strike the right balance between recall and precision in domains with different risk tolerances. Those questions are not resolved by the current deployment, but they delimit the research agenda around organizationally embedded LLM triage.

7. Nomenclature and related usages

In the arXiv record, “RescueLens” primarily denotes the food-rescue feedback system described above (Raman et al., 19 Nov 2025). However, the supplied literature also contains a different usage in which the details of “LensID: A CNN-RNN-Based Framework Towards Lens Irregularity Detection in Cataract Surgery Videos” are explicitly “recast here as ‘RescueLens’” for a surgical video pipeline (Ghamsarian et al., 2021). That recast framework concerns phase recognition, semantic segmentation, and lens-irregularity statistics in cataract surgery videos rather than volunteer feedback or food rescue operations.

This naming overlap should not be interpreted as a shared technical lineage. The food-rescue RescueLens is an LLM-powered system for categorization, ranking, and direction rewriting, while the recast surgical “RescueLens” is a CNN–RNN and segmentation pipeline defined around VGG19, ResNet50, AdaptNet, and post-processing of lens and pupil masks. The overlap is nominal rather than methodological.

Within the food-rescue context, the term is therefore best understood as denoting a production-deployed, prompt-based triage system for volunteer feedback. Any broader usage requires explicit disambiguation by domain and citation.