FaMA: AI Assistant for Facebook Marketplace

• Facebook Marketplace Assistant (FaMA) is an LLM-powered tool that transforms complex GUI workflows into conversational interactions for streamlined seller and buyer operations.
• It employs ReAct prompting with structured memory and explicit user confirmations to ensure high task success rates and operational transparency.
• Experiments report a 98% success rate and up to 2x speed improvements, demonstrating FaMA's robust performance in automating consumer-to-consumer marketplace tasks.

Facebook Marketplace Assistant (FaMA) is an LLM-powered, agentic assistant for consumer-to-consumer marketplace workflows that shifts interaction from complex GUI sequences to a conversational, goal-oriented interface. It interprets natural-language commands, automates seller-side operations such as listing renewal, listing updates, and bulk messaging, and supports buyer-side conversational product discovery through tool use, structured memory, and explicit action confirmation. Reported experiments show a 98% task success rate on complex Marketplace tasks and up to a 2x speedup on interaction time, positioning FaMA as a lightweight and more accessible alternative to traditional app interfaces (Yan et al., 4 Sep 2025).

1. Problem setting and functional scope

Consumer-to-consumer marketplaces such as Facebook Marketplace, eBay, and Etsy entail high-friction workflows within complex GUIs. Sellers face repetitive, multi-step listing management, including creating, renewing, relisting, and editing, alongside high-volume, duplicative messaging. Buyers confront discovery across vast, unstructured inventory using rigid filters and manual keyword trial-and-error. These tasks are time-consuming and error-prone, increase cognitive load, and limit accessibility (Yan et al., 4 Sep 2025).

FaMA addresses this setting by introducing an LLM-powered, agentic AI that turns natural-language instructions into validated actions. Its seller-side scope includes simplified updating and renewal of listings and the ability to send bulk messages. Its buyer-side scope centers on conversational search. The implemented interaction model is deliberately conservative: every state-changing action is surfaced to the user and executed only after explicit confirmation. This design couples automation with stepwise transparency rather than autonomous execution.

The assistant also supports text and voice inputs. Voice is transcribed upstream by an Automatic Speech Recognition component into text, after which the core system operates as a text-based agent. The resulting entry point is conversational, but the task domain remains operational rather than open-ended: listing operations, search, messaging, and grounded policy answers define the primary action space.

2. Agent architecture, memory, and control loop

FaMA’s architecture comprises a conversational interface, an LLM core, structured memory, a tools and action layer, execution and actuation logic, and guardrails. The reasoning engine is Llama-4-Maverick-17B-128E-Instruct. A system prompt defines the persona as an AI assistant for a Facebook Marketplace user, specifies tool descriptions, and constrains response formatting. ReAct prompting with Chain-of-Thought organizes processing into explicit Thought → Action → Observation cycles (Yan et al., 4 Sep 2025).

Its memory subsystem has three distinct components. The scratchpad is structured short-term memory storing chronological Thought–Action–Observation triplets for multi-step context preservation. Dialog history is a session-scoped, ephemeral transcript of user and agent turns and is purged after inactivity for privacy and context hygiene. Listings Information Memory injects the user’s listing metadata—title, description, and ID—into the session, enabling resolution of natural-language references such as requests to modify a specific listing without manually supplying an identifier.

The tool layer includes Listing Operation Tools for creating, updating, and renewing listings; Inventory Search Tools for querying Marketplace search APIs with constraints parsed from user language; Messaging Tools for single and bulk messages; and a RAG tool that retrieves help and policy content from a vector database of help articles. Tool calls are proposed by the LLM, then executed only after user confirmation. Observations returned by the tools are appended to the scratchpad and fed back into the next reasoning step.

The loop is formalized with agent state $s_t$, user input $u_t$, proposed action $a_t$, and tool observation $o_t$. Planning is written as $a_t = \pi(s_t, u_t)$, where $\pi$ is the LLM policy instantiated by ReAct prompting. If the user confirms, $o_t = \mathrm{Env}(a_t)$ and the state is updated to $s_{t+1} = \mathrm{update}(s_t, u_t, a_t, o_t)$. The objective is to maximize task success under a step limit while minimizing the number of steps for optimality. The system departs from autonomous ReAct by enforcing single-step interactive mode, which pauses after each proposed action for confirmation and functions as both a safety mechanism and a UX transparency mechanism.

3. Retrieval and search substrate

A broader technical realization of FaMA’s buyer-facing search stack can leverage Aug2Search, an embedding-based retrieval framework that uses LLM-generated synthetic data to compensate for sparse, noisy, and biased Marketplace logs. The motivating claim is that real Marketplace search logs are de-identified and aggregated, sparse in query diversity and long-tail patterns, contain low-quality seller-authored listing text, and reflect historical bias; engagement signals such as Click and Listing Interactions cover only a slice of the full query-product intent space. These properties limit the ability of Embedding-Based Retrieval models to learn robust semantic matching (Xi et al., 21 May 2025).

Aug2Search models query-product relevance with a multimodal, multitask two-tower architecture in a shared semantic space. The query tower encodes raw query text with a two-layer XLM encoder, hashes character trigrams into embeddings via EmbeddingBag with sum pooling, encodes country via EmbeddingBag, and produces a final query embedding through attention-weighted fusion of text, character, and country signals. The document tower combines title and description text, character trigrams, context features such as price, category, and creation time, and pre-trained vision embeddings from GrokNet. Text, image, and context are merged by a transformer-fusion encoder initialized from a six-layer XLM-R; the final product embedding is the projected $[\mathrm{CLS}]$ token at the last fusion layer, with attention-weighted fusion across modalities.

Synthetic data generation follows three strategies: S1 generate queries from original listings, S2 enhance product listings, and S3 enhance listings and then generate queries from the enhanced text. The pipeline uses coherence and relevance checks, hallucination minimization, Distinct-2 diversity measurement, BM25-L retrieval gating with MRR@K, human evaluation, and commerce policy alignment under the 3H criteria of helpful, honest, and harmless. Eight Llama variants were studied for generation, but Llama3-8b-instruct was selected for mass generation because M8 and M5 produced excellent quality while M2 offered strong performance and better QPS. The final T2-Detailed prompt outperformed the simpler T1-Basic template in coherence, conversational quality, and relevance.

The training objective is explicitly multitask:

$$L(\theta) = \lambda_1 L_{\mathrm{relevance}} + \lambda_2 L_{\mathrm{engagement}},$$

with $u_t$0 and $u_t$1. The relevance term uses a scaled softmax over in-batch cosine similarities with $u_t$2, and the engagement term uses binary cross-entropy on hard negatives. Training uses Adam, batch size 512, learning rate $u_t$3 for the XLM encoder and $u_t$4 for other components, dropout 0.1, gradient clipping 1.0, 2000 warmup steps, and ROC_AUC-based early stopping.

Empirically, the strongest synthetic strategy is S3. With 100M data points, ROC_AUC improved from 0.600 on original data to 0.640 with S3; at 100M, S3 reached PBC-o 0.395, PBC-r 0.360, RCR 0.840, and ROC_AUC 0.640. In mixed-data training at 125M total examples, 50M original + 100M synthetic, S3 reached ROC_AUC 0.660. The reported result that synthetic-only models often outperform original-only or mixed models at the same data volume suggests that FaMA’s retrieval component can be trained to generalize beyond the historically observed query-product space. For deployment, the proposed stack includes product vector indices such as HNSW or IVF-PQ, daily offline indexing with streaming updates for new listings, multilingual support through XLM/XLM-R encoders and locale features, and latency optimizations via precomputed product embeddings and hot-index caching.

4. Listing generation and multimodal content assistance

FaMA’s seller-side assistance can also be understood through related work on multimodal listing generation. IPL is an Intelligent Product Listing system designed for C2C platforms in which users compose product descriptions by uploading photos of the selling product. The underlying rationale is specific to C2C commerce: unlike B2C listings, C2C descriptions must be colloquial, authentic, and trust-building, and are written by individual sellers who often lack e-commerce expertise (Chen et al., 2024).

IPL uses Qwen-VL (7B) as its base MLLM, with the visual encoder frozen while the VL-Adapter and LLM components are updated. Category prediction is performed with ALBEF, reporting >80% accuracy across tens of thousands of categories. Visual retrieval identifies identical or similar products via vector search with thresholding; reported offline accuracy is >60% for identical-product retrieval and >90% for similar-product retrieval. Retrieved listings supply structured attribute values, extracted by the domain-specific MLLM into JSON according to category-specific attribute templates. Generation then conditions on images, predicted category, templates, and retrieved reference attributes.

The core training objective is next-token generation,

$u_t$5

with loss computed on model output tokens and prompt prefixes excluded. Training uses 1.27M mixed-modality instructions, 3 epochs, 4,788 training steps, global batch size 768, BF16, AdamW, and 24 A800 GPUs for approximately 16 hours. Inference optimizations include quantization, ViT operation optimizations, key-value caching, kernel fusion, parallel computation, and streaming output, reducing average response time from ~5s to <3s.

The strongest technical result concerns multimodal RAG. In ablation, Image only achieved human attribute accuracy 0.36, while Image + Reference reached 0.74, and Image + Category + Reference reached 0.75. This indicates that retrieved reference attributes rather than category labels are the primary driver of factuality. Online deployment results are also notable: 72% of users published listings based on generated content, 32% adopted more than 50% of the generated text, and AI-assisted listings had a 5.6% higher quality score than those without assistance. For FaMA, this line of work suggests a concrete multimodal listing assistant that begins with photos, retrieves comparable items, constrains generation through category-specific templates, and reduces hallucination by grounding outputs in retrieved evidence.

5. Query categorization and personalized retrieval

FaMA’s conversational interface depends on a stronger query-understanding substrate than raw lexical matching. HierCat, the deployed query categorization system for Facebook Marketplace, addresses the vagueness of Marketplace queries, the noise of weak supervision, and multilingual data imbalance through a dual-encoder architecture and a hierarchical inference step over the Facebook Product Taxonomy. The taxonomy has around 6,000 categories and a maximum depth of 6, and the system explicitly allows a correct prediction to terminate at an internal node rather than forcing a leaf label (He et al., 2023).

The query encoder is a 2-layer XLM transformer fused with a character trigram EmbeddingBag via attention. The category encoder is an EmbeddingBag over category text. Both towers produce 128-dimensional embeddings and are scored by cosine similarity. Training data consist of 24M weakly supervised $u_t$6query, category$u_t$7 pairs mined from 14 days of de-identified and aggregated engagement logs. Hierarchical inference propagates child probabilities upward and renormalizes scores level by level, producing path-consistent category predictions. Retrieval pretraining from an embedding-based retrieval model improves fine-grained categorization, with the reported ablation showing +3% F1 at L3 and +16% F1 at L6, together with +4% acc@5 at L6, while remaining neutral at L1.

Online impact is reported directly. Structured retrieval boosts based on L1 and L2 category predictions produced a +1.4% improvement in NDCG, and adding L3 boosts yielded a +4.3% increase in searcher engagement. In a FaMA setting, these predictions can support category-aware suggestions, automatic filter proposals, and ambiguity management when the user query is coarse, underspecified, or multilingual.

Personalized retrieval introduces a second, older but still relevant substrate. A collaborative multimodal retrieval system for Facebook Marketplace uses a dual-encoder architecture over users and products, trained on buyer-initiated messages. The item tower combines TextCNN representations of title and description with ResNet-50 image features; the user tower aggregates keyword embeddings and demographics. Both project into a 50-dimensional L2-normalized space and are trained with a pairwise margin-based ranking loss with margin $u_t$8 (Zheng et al., 2018).

The operational motivation is explicit: Facebook Marketplace must score tens of thousands of products within a couple hundred milliseconds while handling acute cold start because listings are short-lived and interaction data are sparse. The reported online result is a +26.95% increase in buyer-initiated messages versus a time-based retrieval baseline, and the abstract reports a +9.58% improvement in offline prediction accuracy. Within FaMA, this kind of retrieval layer complements HierCat: HierCat supplies category-consistent intent structure, while collaborative multimodal retrieval supplies personalized candidate generation under strict latency and cold-start constraints.

6. Evaluation, safety, privacy, and limitations

FaMA was evaluated through automated simulation and timing studies. The automated environment used an LLM-based user simulator interacting with FaMA; both were powered by Llama-4-Maverick-17B-128E-Instruct to isolate architectural merits. The dataset contained 100 synthetic marketplace listings spanning electronics, vehicles, toys, and furniture. Tasks included Inventory Search, Renew Listing, and Bulk Reply, and each task was subject to a 5-step interactive limit. Task Success Rate and Task Optimality Rate were defined as

$u_t$9

The timing study compared FaMA against the standard Marketplace mobile app for Bulk Messaging Reply and Filtered Inventory Search (Yan et al., 4 Sep 2025).

Task	Reported performance	Timing
Inventory Search	98% success; among successful, 100% optimal single-step completion	FaMA 15 sec vs. Manual 25 sec
Renew Listing	Text notes 100% success	Minimal baseline: 2 steps
Bulk Reply	Success rate over 96%; among successful, >84% optimality	FaMA 25 sec vs. Manual 50 sec

These results indicate high reliability under the study protocol, but the scope is bounded. The paper does not report ablation studies or detailed failure-mode analysis. The overall failure rate across complex tasks remained below 2%, yet the evaluation used synthetic listings and a simulator rather than live user traffic. The timing study involved authors as experienced Marketplace users, and broader user studies and statistical testing such as confidence intervals were not reported.

Safety and privacy are built into the architecture rather than added post hoc. State-changing actions such as listing edits, renewals, and message sending require confirmation at every step. Dialog history is session-scoped and purged after inactivity. The RAG tool confines policy and how-to answers to a vetted knowledge base of help articles. At the same time, operational monitoring details, content filters, rate limits, and recovery strategies for API failures are not discussed beyond these safeguards.

Related retrieval work introduces additional system-level risks. Synthetic-data-driven retrieval can suffer from domain shift if synthetic data reflect LLM priors more than real user behavior; it can encode socioeconomic or brand biases; and it can overfit to synthetic intents if not calibrated against fresh engagement data. These risks motivate fairness audits, locale balancing, factuality checks, and continued interleaving of real engagement signals in FaMA’s training and monitoring loop (Xi et al., 21 May 2025).

7. Future directions and broader significance

Several future directions follow directly from the reported designs. For the core agent, the immediate agenda includes more comprehensive real-user studies, explicit ablation analyses of planning, memory, and tool contributions, expanded safety and monitoring policies, and multimodal extensions for image understanding, since the implemented FaMA architecture supports voice through ASR but otherwise operates with a text-only LLM core (Yan et al., 4 Sep 2025).

For the retrieval layer, the proposed extensions are more expansive. Instruction-tuned LLMs can be used for improved intent modeling and associative query generation; reinforcement learning from marketplace feedback can optimize retrieval against multi-objective signals such as CTR, message rates, and conversion proxies; on-device retrieval with quantized query towers can support mobile assistants; multimodal pretraining can strengthen image-text alignment in listing embeddings; and generative retrieval combined with RAG and knowledge graphs can support query rewriting and attribute normalization. List-aware reranking and truncation strategies are also proposed as a way to jointly optimize retrieval and downstream ranking (Xi et al., 21 May 2025).

Taken together, these lines of work suggest that FaMA is best understood not as a single chatbot, but as a layered marketplace system. At the top layer, it is an agentic interface grounded in ReAct-style planning, scratchpad memory, and explicit tool execution. Beneath that interface, it depends on query categorization, multimodal retrieval, synthetic data augmentation, and grounded listing generation to make conversational interaction operationally meaningful. The significance of FaMA therefore lies less in replacing search or listing infrastructure than in reconfiguring them behind a conversational control surface that can execute, explain, and validate Marketplace actions in a unified interaction model.